Skip to content

Official Codebase of 3DCodeVerse

Python

9b: one hour of self-generated DPO lifts four dialects, OpenSCAD 6% to 90%

gaoypenge3f14b6ba381
diff --git a/finetune/configs/lf/export_9b_mmvis.yaml b/finetune/configs/lf/export_9b_mmvis.yaml
new file mode 100644
index 00000000..665a1dfd
--- /dev/null
+++ b/finetune/configs/lf/export_9b_mmvis.yaml
@@ -0,0 +1,12 @@
+### Merge a LoRA adapter into the base model (LLaMA-Factory): llamafactory-cli export <this file>
+model_name_or_path: /wekafs/ict/hx_624/models/Qwen3.5-9B
+adapter_name_or_path: /wekafs/ict/hx_624/llm-ft/runs/lf_qwen35_9b_mmvis
+template: qwen3_5
+enable_thinking: false
+trust_remote_code: true
+finetuning_type: lora
+
+export_dir: /wekafs/ict/hx_624/llm-ft/runs/lf_qwen35_9b_mmvis/merged
+export_size: 5
+export_device: cpu
+export_legacy_format: false
diff --git a/finetune/configs/lf/export_stopdpo.yaml b/finetune/configs/lf/export_stopdpo.yaml
new file mode 100644
index 00000000..637cc067
--- /dev/null
+++ b/finetune/configs/lf/export_stopdpo.yaml
@@ -0,0 +1,12 @@
+### Merge a LoRA adapter into the base model (LLaMA-Factory): llamafactory-cli export <this file>
+model_name_or_path: /wekafs/ict/hx_624/llm-ft/runs/lf_qwen35_9b_mmmix2/merged
+adapter_name_or_path: /wekafs/ict/hx_624/llm-ft/runs/lf_qwen35_9b_stopdpo
+template: qwen3_5
+enable_thinking: false
+trust_remote_code: true
+finetuning_type: lora
+
+export_dir: /wekafs/ict/hx_624/llm-ft/runs/lf_qwen35_9b_stopdpo/merged
+export_size: 5
+export_device: cpu
+export_legacy_format: false
diff --git a/finetune/configs/lf/lora_9b_full.yaml b/finetune/configs/lf/lora_9b_full.yaml
new file mode 100644
index 00000000..cbfd89c7
--- /dev/null
+++ b/finetune/configs/lf/lora_9b_full.yaml
@@ -0,0 +1,44 @@
+model_name_or_path: /wekafs/ict/hx_624/models/Qwen3.5-9B
+trust_remote_code: true
+flash_attn: fa2
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_rank: 64
+lora_alpha: 128
+lora_dropout: 0.05
+lora_target: all
+freeze_vision_tower: true
+freeze_multi_modal_projector: true
+dataset_dir: /wekafs/ict/hx_624/llm-ft/data/full9b
+dataset: 3dcv_mm_img_text_train,3dcv_mm_img_train,md_full9b_train
+eval_dataset: 3dcv_mm_img_text_val
+template: qwen3_5
+enable_thinking: false
+cutoff_len: 8192
+preprocessing_num_workers: 16
+dataloader_num_workers: 16
+neat_packing: false
+packing: false
+tokenized_path: /wekafs/ict/hx_624/llm-ft/data/lf/tok_full9b_len8192
+output_dir: /wekafs/ict/hx_624/llm-ft/runs/lf_qwen35_9b_full
+logging_steps: 5
+save_steps: 600
+save_only_model: true
+plot_loss: true
+overwrite_output_dir: true
+report_to: tensorboard
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 4
+learning_rate: 0.0001
+num_train_epochs: 1.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.05
+bf16: true
+gradient_checkpointing: true
+ddp_timeout: 180000000
+resume_from_checkpoint: null
+per_device_eval_batch_size: 1
+eval_strategy: steps
+eval_steps: 600
+save_total_limit: 4
diff --git a/finetune/docs/REPORT.md b/finetune/docs/REPORT.md
new file mode 100644
index 00000000..61cd505e
--- /dev/null
+++ b/finetune/docs/REPORT.md
@@ -0,0 +1,879 @@
+# 3D-code LLM finetune 调研报告(dgx03 / 4×H100,2026-08-21 → 08-22)
+
+> 目标:在前 4 张 H100 上跑通「文本 → Blender Python 3D 代码」的 LLM 微调流程,评估 3DCodeVerse 数据是否适合训练,为后续 large-scale finetune 给出配置与结论。
+> 工程目录:`/wekafs/ict/hx_624/llm-ft`(README 有命令)。**所有数字见第 5 / 7 / 8 节;最近更新 2026-08-28 12:00(新增 §12:全量数据清洗与编译/渲染验证、URDF 恢复、泄漏探针、图像条件生成);上一版 2026-08-25 20:30(§11 完成:评估方法论、md_max 审计、27B 零样本/v1/v2/v3/DPO、scipy 修正、caption 增强对照;**27B v2 = 3DCodeBench 93.4% + 最佳几何**,9B+执行反馈 DPO = 96.2%;完整实验清单见 `llm_finetune_exps.md`,数据已发布到 HF `ilabai/3dcodeverse-llamafactory`)。**
+
+## 1. 环境(已 setup,可直接复用)
+| 项 | 内容 |
+|---|---|
+| 机器 | dgx03,8×H100 80GB(只用 GPU 0–3;GPU 3 长期被 pxiang 占用,GPU 2 也会被其推理任务临时抢占 → 实际稳定可用的只有 0/1,**多卡实验前需要和 pxiang 协调**),driver 570 / CUDA 12.8,224 核,2 TB RAM |
+| 存储 | `$HOME` 已满(~450 MB 可用)→ 一切装在 `/wekafs/ict/hx_624`:Anaconda 2026.07 → `anaconda3/`,HF cache → `cache/huggingface`,模型 → `models/`,数据 → `data/` |
+| conda env `llmft` | torch 2.9.1+cu128, transformers 5.15.1, trl 1.10, peft 0.20, flash-attn 2.8.3, fla 0.5.2, deepspeed 0.19.5 —— 用于数据处理、HF 推理评测 |
+| conda env `lf` | **LLaMA-Factory 0.9.6.dev0**(`tools/LLaMA-Factory`)+ torch 2.8.0+cu128 + transformers 5.8.0 + flash-attn 2.8.3 + fla 0.5.2 + causal-conv1d 1.7.0 + **triton 3.7.1** + deepspeed 0.19.5 —— 用于训练 |
+| Blender | 5.0.1 headless(`tools/blender-5.0.1-linux-x64/blender`),用于执行生成代码并导出 GLB |
+
+踩过的坑(后续 large-scale 训练要注意):
+1. LLaMA-Factory 拒绝 torch 2.9.x + 含 Conv3D 的模型(Qwen3.5 有视觉塔)→ `lf` 环境用 torch 2.8.0(或 ≥2.10)。
+2. Qwen3.5 的线性注意力层(Gated DeltaNet)靠 `flash-linear-attention`(fla, Triton) 加速;Hopper 上 **Triton 3.4–3.7.0 的反向内核有 bug**,fla 直接拒绝运行 → 必须 `pip install triton==3.7.1`(已验证与 fp32 参考实现一致,相对误差 <0.5%)。没有 fla/causal-conv1d 时 transformers 退回纯 torch 实现(慢很多)。
+3. wekafs 上 pip/conda 解包慢(Anaconda 解包 ~10 min),HF 下载用 `hf download`(xet)比单线程 curl 快很多;hf 的文件锁在 wekafs 上有时报 "waiting to acquire lock",小文件用自己写的 `scripts/fetch_hf_files.py` 更稳。
+4. torchaudio 版本不匹配会把 transformers import 搞挂(`llmft` 里卸掉了)。
+5. LLaMA-Factory `neat_packing: true` 要求 **eval 的 batch size = 1**(`per_device_eval_batch_size: 1`),否则第一次 eval 时 DataLoader 断言崩溃;配合 `save_steps` 要小于 `eval_steps`,否则像我第一次那样 30 min 训练在第 100 步崩掉且没有任何 checkpoint。
+6. vLLM 的 torch.compile 缓存默认写 `~/.cache/vllm`(home 已满 → "No space left on device" 直接起不来)→ 统一 `export VLLM_CACHE_ROOT=/wekafs/ict/hx_624/cache/vllm`(已写进 `env.sh` 与评测脚本)。
+7. Triton 内核缓存(`TRITON_CACHE_DIR`)放在 wekafs 上时,多进程同时编译同一内核偶发 `FileNotFoundError ...chunk_gated_delta_rule_*.json`(原子 rename 在网络盘上不可靠)→ 同机并行跑多个训练时会反复出现;已改为 **每个训练 run 用本地盘上独立的 Triton 缓存**(`lf_train.sh` 自动从共享缓存复制一份到 `/tmp/hx_624_triton/<run>`),之后不再出现。
+8. DPO 的显存是 SFT 的 ~2 倍以上(chosen+rejected 拼接前向 + 参考模型):8k cutoff 下 GLSL/OpenSCAD 的长偏好对在 80 GB 卡上 OOM → 多语言 DPO 用 `cutoff_len 4096`(或过滤长对)。
+9b. **Blender 自带的 Python 里要装 scipy**(`5.0/python/bin/python3.11 -m pip install scipy`):3DCodeBench 的 212 个参考脚本里有 15 个 import scipy,缺了它会把"写法正确"的生成判为失败(详见 §11.6,这让 27B 微调版凭空少了 8.1 个点)。
+9. Blender 执行器:不要在脚本里调 `read_factory_settings()`(context 会过期);只统计/导出 view layer 里的 mesh;生成脚本里所有路径必须用绝对路径(runner 会 chdir)。
+
+## 2. 模型选择
+| 候选 | 说明 | 结论 |
+|---|---|---|
+| **Qwen/Qwen3.5-9B**(2026-03) | 10B 以下最新最热的官方稠密模型(HF 月下载 13.7M),hybrid Gated-DeltaNet + full attention,原生 256K 上下文,thinking 可关 | **主选**;LLaMA-Factory 模板 `qwen3_5` + `enable_thinking: false` |
+| Qwen/Qwen3.5-4B | 同架构更小 | 备选(更快,单卡即可全参) |
+| Qwen/Qwen3-8B(2025-04) | 标准稠密 transformer,零工程摩擦 | 备选(如 fla/triton 再出问题) |
+| Qwen/Qwen2.5-Coder-7B-Instruct(2024-09) | 老牌 coder,已下载并跑了基线 | 基线参考 |
+| Qwen3-Coder-30B-A3B / Qwen3-Coder-Next / KAT-Coder-V2.5 | MoE,> 10B 总参 | 不符合 ≤10B 要求;LoRA 可行但不在本次范围 |
+
+## 3. 数据:3DCodeVerse 及相关数据集
+### 3.1 格式结论
+* `ilabai/3dcodeverse` 每个子集都是 **`metadata.parquet`(含 `code`、`captions{detailed,instruction,factory}`、`meta_json`)+ `samples-*.tar`(渲染图/GLB)**。**做 LLM 文本训练只需要 parquet(全部 173 个文件 <1 GB)**,无需下载 285 GB 的 tar。→ `data/3dcodeverse_meta/`
+* captions 三种字段全都齐(≥99.8%):`instruction`(任务式指令,最适合做 user prompt)、`detailed`(外观描述)、`factory`(代码做了什么的描述)。
+* 例外:`articraft/urdf_*` 的 `code` 列为空(多文件 URDF 工程,需从 tar 取);`shadertoy` 是多文件(`shader_json` 列 + `code.glsl`)。
+* `ilabai/thingiverse-openscad-blender-5-0`(1193)与 `...-5-1-2-parametric`(6919)是**按样本目录存放**(code.py / prompt.txt / captions.json / openscad/*.scad)→ 需要自己抓小文件拼 pair。parametric 版的 `code.py` 每个 ~3000 行,其中 ~2500 行是 OpenSCAD→bpy 的 shim 样板,**不适合直接当 SFT 目标**;它的 `prompt.txt → *.scad` 倒是干净的 OpenSCAD 指令对。5-0 版里 166 个是 `reference_mesh.json` 回退(读 json 重建 mesh),要过滤掉。
+* `3dcodebench/3dcodebench`(212 任务)= 3dcodeverse `3dcodebench/factories_*` 的子集(243 个 factory 里 212 个同名)→ **训练里必须剔除这 212 个 factory,否则测试集泄漏**。
+
+### 3.2 规模(代码 token 用 Qwen3.5 tokenizer 抽样估计)
+| 来源 | 样本数 | 语言 | 代码 token 总量 | 单样本 token p50/p90 | 备注 |
+|---|---|---|---|---|---|
+| deepcad (v1…v1_4) | 128,201 | CadQuery Python | ~84M | 430–510 / ~1.3k | 短、规整,captions 为几何描述 |
+| shadertoy (112 shard) | 119,961 | GLSL | ~167M | 550–1300 / 2.4k–4.5k | 多文件,license CC-BY-NC-SA |
+| articraft cadquery_* | 6,882 | CadQuery (+URDF 关节) | ~71M | 8k–10k / 15k–18k | 每个脚本带大段 shim 样板 |
+| articraft urdf_* | 6,146 | URDF | — | — | parquet 里无 code |
+| bioinspired3d | 6,343 | **Blender Python** | ~2.8M | 390–450 / 570–740 | 短脚本(p50 50 行),22% 重复代码 |
+| 3dcodebench factories | 486 | **Blender Python** | ~5.3M | ~9.2k / ~20k | Infinigen 风格长脚本;212 个与测试集重名 |
+| thingiverse/openscad | 1,193 | OpenSCAD | ~2.5M | 1.1k / 4.5k | 对应 blender-5-0 转换版 1193 个 bpy 脚本 |
+| blender_distill | ~100 | Blender Python | — | ~3k | Claude 生成,质量高但量少 |
+| **合计** | **~270k** | | **~330M tokens** | | |
+
+### 3.3 对本次训练的取舍(sft_v1)
+Blender-Python 目标(对齐 3DCodeBench):bioinspired3d(去重)+ 31 个非测试 factory + blender_distill + thingiverse-5-0 原生 CSG 子集 → 去重、丢掉 >8192 token 的样本后 **5,193 train / 105 val,约 3.8M token/epoch**。prompt 随机取 `instruction`(60%) / `detailed`+包装句(30%) / 原始 name(10%),system prompt 固定,assistant 输出 ```` ```python ```` 代码块。
+问题:训练集代码偏短(p50 ≈ 600 token),而测试集参考代码 p50 ≈ 360 行 / 数千 token,分布差异大 → 见第 6 节建议。
+**数据构建过程中发现的问题(重要):**
+* `thingiverse-openscad-blender-5-0` 全部 1193 个 `code.py` 都内嵌 `load_mesh_fallback()`(读取 `assets/reference_mesh.json`),连 405 个标记为 `native_csg_generated` 的样本也带这个回退分支 → 我的过滤器把它们全部剔掉了,**sft_v1 里实际上没有 thingiverse 样本**。进一步看,这 405 个"原生"脚本其实是**通用 CSG 解释器 + 外部 `assets/expanded_csg.json`**(`CSG_TREE = json.loads(...)`; `build(CSG_TREE)`),几何逻辑不在 code.py 里 → **5-0 版本两类样本都不能直接当 text→code 训练数据**。`...-5-1-2-parametric` 版才是源码级翻译(逻辑在 code.py 里),但每个文件前 ~2500 行是 OpenSCAD→bpy 的 shim;要用它训练,应把 shim 抽成可 import 的库、只保留样本特有的 ~500 行作为训练目标(或者直接用它的 `prompt.txt → *.scad` 做 OpenSCAD 方言的 SFT)。
+* 3dcodebench factories 脚本 p50≈9.2k token、p90≈20k token:在 cutoff 8192 下 **60% 被丢弃**(v2_indist 只保留了 212 个测试 factory 中的 168 个较短样本)。要学 Infinigen 风格的长脚本,cutoff 至少 16k–24k(显存 OK:LoRA 单卡 bs1+grad ckpt 可到 32k)。
+* bioinspired3d 有 22% 代码重复(同一脚本配不同 prompt),去重后 geo 4557→3563;tex 子集和 geo 子集是同一来源的配对(tex 1786 全部 ⊂ geo key)。
+* 数据变体规模:v1 5193 / v2_indist 5362 / v3_bio 4998 / v4_cq 11069(+5877 deepcad) / v5,v6 5193(同 v1,仅 prompt 字段不同)。
+
+## 4. 训练配置(LLaMA-Factory,均已写在 `configs/lf/`)
+(数字见第 5 节)
+* **LoRA**:`qwen35_9b_lora_sft.yaml` — rank 64 / alpha 128 / all linear,视觉塔冻结,cutoff 8192,packing+neat_packing,bf16,grad ckpt,lr 1e-4 cosine,2 epoch。
+* **全参**:`qwen35_9b_full_sft.yaml` — DeepSpeed ZeRO-3,lr 1e-5,其余同上。
+* 预处理一次:`qwen35_9b_preprocess.yaml` → `tokenized_path`,两种训练复用。
+* 启动:`GPUS=0,1,2 scripts/lf_train.sh configs/lf/<yaml>`;LoRA 合并:`scripts/lf_export.sh configs/lf/export_lora_template.yaml`。
+
+## 5. 实验矩阵与结果
+评测协议(`eval/`):212 个 3DCodeBench prompt(`prompt_instruction.txt`)→ 模型生成(greedy,最多 6144 新 token,Qwen3.5 关 thinking)→ 抽取 ```python 块 → Blender 5.0.1 无头执行 + 导出 GLB(超时 300 s)→ 与 GT GLB 比较(归一化到单位包围盒,采样 1 万点,Chamfer 距离与 [email protected]/0.1,取 4 个绕竖轴旋转里最好的)。参考脚本自检:Chamfer≈0.016、[email protected]≈1.0。
+
+| 实验 | 模型 | 训练数据 | 方法 | 说明 |
+|---|---|---|---|---|
+| base-coder7b | Qwen2.5-Coder-7B-Instruct | – | zero-shot | 老 coder 基线 |
+| base-q35 | Qwen3.5-9B (non-thinking) | – | zero-shot | 新模型基线 |
+| lora-v1 | Qwen3.5-9B | sft_v1(bio3d+31 factory+distill+thingi50, 5.2k) | LoRA r64, 2 ep | 主实验 |
+| lora-v2_indist | Qwen3.5-9B | v1 + 212 个与测试集同名 factory(geo+tex) | LoRA | **泄漏/分布内上界**:看 Infinigen 风格长代码的作用 |
+| lora-v3_bio | Qwen3.5-9B | 只用 bioinspired3d | LoRA | 短脚本数据够不够 |
+| lora-v4_cq | Qwen3.5-9B | v1 + 6k deepcad CadQuery | LoRA | 跨方言数据是否有益 |
+| lora-v5_instr / v6_detail | Qwen3.5-9B | v1,prompt 只用 instruction / 只用 detailed | LoRA | prompt 字段消融 |
+| lora-v2b_indist16k | Qwen3.5-9B | v2 同源,cutoff 16k(纳入 347 条测试同名 factory) | LoRA | 长脚本是否被截断的影响 |
+| lora-v7_ep4 / v8_r16 / v1_rep | Qwen3.5-9B | v1 | LoRA 4 epoch / rank 16 / 换 seed 并用 2 卡 DDP 复跑 | epoch、rank、随机性与多卡扩展 |
+| ckpt sweeps | – | v1 @step100,v3_bio @step50/100 | 中间 checkpoint 评测 | 数据效率曲线 |
+| lora-v10_4b | **Qwen3.5-4B** | v1 | LoRA(同配置) | 小一档模型够不够 |
+| lora-v9_lr2e4 | Qwen3.5-9B | v1 | LoRA lr 2e-4 | 学习率消融 |
+| lora-v12_qwen3_8b | **Qwen3-8B**(2025 标准稠密) | v1 | LoRA(同配置,模板 qwen3) | 新旧模型对比:Qwen3.5 是否真的更好 |
+| v11b_selfprompts | Qwen3.5-9B | v1 + 8.3k 条(训练 prompt 上采样、执行通过、去重) | 执行验证自训练 | 同 prompt 多解是否有用 |
+| v13_boot | Qwen3.5-9B | v1 + 15.7k 条(**新 prompt** 上采样、执行通过、去重) | 执行验证自举 | 新 prompt 自举数据是否有用(数据扩 4 倍) |
+| full-v1 | Qwen3.5-9B | sft_v1 | 全参 + ZeRO-3(2×H100 + CPU offload) | LoRA vs 全参;多卡配置/吞吐。因 GPU 2 被占且 2 卡 offload 太慢(65 s/step),只训到 1 epoch(step 60)评测 |
+| full-v2_3gpu | Qwen3.5-9B | sft_v1 | 全参 + ZeRO-3 **3×H100(GPU 0,1,3)无 offload**,2 epoch | 正式的多卡全参配置与吞吐(GPU3 空出后补做) |
+| dpo_exec_v1 | LoRA-v1 合并模型 + 新 LoRA | 2590 对「同 prompt 下执行通过 vs 报错」的偏好对(来自两轮采样) | **DPO(执行反馈偏好优化)** | 执行反馈能否直接压低错误率 |
+| dpo_exec_v2 | DPO_v1 合并模型 + 新 LoRA | DPO_v1 在 9.9k 新 prompt 上采样 2 次 → Blender 过滤 → 新的 (通过, 报错) 对 | **迭代 DPO 第二轮** | 迭代执行反馈是否继续提升 |
+| dpo_md_mixed | md_mixed 合并模型 + 新 LoRA | md_mixed 在每种语言 1.5k 训练 prompt 上采样 2 次 → 各语言执行器过滤 → (通过, 报错) 对 | **多语言执行反馈 DPO** | 执行反馈 DPO 是否能推广到四种语言 |
+| **md_big** | Qwen3.5-9B | 多语言大集:Blender 5.0k + 15.7k 自举、CadQuery 40k、GLSL 40k、OpenSCAD 1.1k(**101.9k 样本、107M token/epoch**,md_mixed 的 4 倍) | LoRA r64,2 epoch,**4 卡 DDP 3.4 h**,每 220 步存 checkpoint | **更多训练量/时间是否带来更明显提升**(训练量曲线)→ 见 §8:3DCodeBench 67% → **84.9%** |
+| bon-dpo | DPO_v1 合并模型 | – | T=0.7 采样 4 次 + 执行筛选 | DPO + best-of-n 的上限 |
+| bon-4b | Qwen3.5-4B LoRA 合并模型 | – | T=0.7 采样 4 次 + 执行筛选 | 4B + best-of-n 能否追上 9B |
+| bon-v1 | LoRA-v1 合并模型 | – | T=0.7 采样 4 次 + Blender 执行筛选 | 看 best-of-n / 执行反馈的上限 |
+
+GPU 利用率说明:HF `generate` 做推理评测时 nvidia-smi 只显示 30–50%(逐 token 解码、带宽受限),这是正常的;LLaMA-Factory 训练(packing + flash-attn)时 GPU 0 稳定在 ~93%。
+
+### 5.1 结果表
+(完整机器可读表:`eval/all_results.json`;exec = 212 题中 Blender 成功执行并导出非空网格的比例;CD/F 只在成功题上算,F(all) 把失败记 0)
+
+| 实验 | exec OK | Chamfer↓(OK) | [email protected](OK) | [email protected](OK) | [email protected](all) | 平均输出 token | 训练时长 |
+|---|---|---|---|---|---|---|---|
+| Qwen2.5-Coder-7B-Instruct(基线) | 18/212 = 8.5% | 0.312 | 0.231 | 0.394 | 0.014 | 1040 | – |
+| Qwen3.5-9B non-thinking(基线) | **0/212** | – | – | – | 0 | 3520 | – |
+| 同上 lenient(修正 `scene.clear`) | 4/212 = 1.9% | 0.265 | 0.239 | 0.486 | 0.004 | – | – |
+| **LoRA-v1**(干净 5.2k 样本,2 ep) | **159/212 = 75.0%** | 0.201 | 0.382 | 0.598 | **0.285** | 774 | 30 min(1×H100) |
+| LoRA-v2_indist(v1 + 168 条测试同名 factory) | 131/212 = 61.8% | **0.186** | **0.411** | **0.642** | 0.252 | 1360 | 37 min(1×H100) |
+| LoRA-v2b_indist16k(v2 同源但 cutoff 16k:347 条测试同名 factory + 更多 distill) | 114/212 = 53.8% | 0.197 | **0.414** | 0.622 | 0.223 | 2580 | 59 min |
+| LoRA-v5_instr(v1 数据,prompt 只用 `instruction`) | 143/212 = 67.5% | 0.199 | 0.403 | 0.609 | 0.270 | 876 | 32 min |
+| LoRA-v6_detail(v1 数据,prompt 只用 `detailed` 外观描述) | 139/212 = 65.6% | 0.208 | 0.369 | 0.588 | 0.240 | 1128 | 34 min |
+| LoRA-v3_bio(只用 bioinspired3d 5.0k) | 156/212 = 73.6% | 0.205 | 0.380 | 0.597 | 0.278 | 539 | 25 min |
+| LoRA-v8_r16(v1 数据,rank 16 / alpha 32) | 162/212 = 76.4% | 0.210 | 0.372 | 0.590 | 0.284 | 667 | 32 min |
+| LoRA-v9_lr2e4(v1 数据,lr 2e-4) | 161/212 = 75.9% | 0.204 | 0.377 | 0.589 | 0.286 | 648 | 32 min |
+| LoRA-v7_ep4(v1 数据,4 epoch) | 147/212 = 69.3% | 0.205 | 0.380 | 0.602 | 0.262 | 857 | 62 min |
+| LoRA-v4_cq(v1 + 5.9k deepcad CadQuery,跨方言) | 157/212 = 74.1% | 0.201 | 0.375 | 0.596 | 0.274 | 694 | 68 min |
+| LoRA-v3_bio @ step 50 / 124(≈0.8 epoch, ≈3M token) | 152/212 = 71.7% | 0.211 | 0.375 | – | 0.267 | – | 12 min |
+| LoRA-v3_bio @ step 100 / 124 | 154/212 = 72.6% | 0.215 | 0.368 | – | 0.264 | – | 24 min |
+| **全参 ZeRO-3**(sft_v1,lr 1e-5,1 epoch = step 60,2×H100 + CPU offload) | 151/212 = 71.2% | 0.200 | 0.381 | 0.604 | 0.269 | 874 | 67 min(2 卡) |
+| **LoRA-v1 best-of-4**(T=0.7 采样 4 次,按 Blender 执行结果选) | pass@4 **193/212 = 91.0%**(单样本平均 70.6%) | – | – | – | **0.462**(取 4 个样本最好) | – | 推理 4×1.4 min |
+| **LoRA-v1 best-of-7**(再加 3 个样本;第 8 个因 GPU 被抢占未生成) | pass@7 **200/212 = 94.3%** | – | – | – | **0.528** | – | 推理 7×1.4 min |
+| LoRA-v1_rep(同 v1 配置,seed=1,**2×H100 DDP**) | 166/212 = 78.3% | 0.196 | 0.393 | 0.607 | 0.302 | 666 | 17.5 min(2 卡)|
+| **Qwen3.5-4B** LoRA(v1 数据,同配置) | 148/212 = 69.8% | 0.202 | 0.378 | 0.614 | 0.262 | 1000 | 24.6 min(1 卡)|
+| **全参 ZeRO-3 3×H100(GPU 0,1,3,无 offload)**,2 epoch | 151/212 = 71.2% | 0.205 | 0.366 | 0.590 | 0.259 | 760 | 52 min*(干净步速 8 s/step ≈ 15 min) |
+| **全参 ZeRO-3 4×H100(GPU 0–3,无 offload,干净)**,2 epoch | 154/212 = 72.6% | 0.207 | 0.377 | 0.585 | 0.272 | – | **11.9 min(4 卡)** |
+| LoRA-v11(v1 + 自训练样本,实际只多 1 条 → 等价于 v1 复跑 #3,2 卡 DDP) | **174/212 = 82.1%** | 0.192 | 0.408 | 0.622 | **0.331** | 617 | 15.6 min(2 卡)|
+| LoRA-v11b(v1 + 8.3k 条训练 prompt 上的执行验证新解) | 171/212 = 80.7% | 0.196 | 0.394 | 0.611 | 0.318 | – | 34 min(2 卡)|
+| LoRA-v13_boot(v1 + 15.7k 条**新 prompt** 自举样本,2 卡) | 163/212 = 76.9% | 0.200 | 0.403 | 0.619 | 0.308 | – | 63 min(2 卡)|
+| **Qwen3-8B**(2025 稠密)LoRA,同配置,模板 qwen3 | 162/212 = 76.4% | 0.220 | 0.364 | 0.560 | 0.277 | 680 | 27.6 min |
+| Qwen3.5-4B LoRA best-of-2(T=0.7,另外 2 个样本因与全参训练抢 GPU 起失败) | pass@2 = 75.0%(单样本 60.4% / 57.1%) | – | – | – | 0.300(best-of) | – | – |
+| (交叉)LoRA-v1 在 **外观描述式** 测试 prompt 上 | 161/212 = 75.9% | 0.197 | 0.387 | 0.607 | 0.294 | – | – |
+| (交叉)LoRA-v6_detail 在外观描述式测试 prompt 上 | 146/212 = 68.9% | 0.207 | 0.381 | 0.588 | 0.262 | – | – |
+| **DPO_exec_v1**(LoRA-v1 + 2590 对执行反馈偏好对 DPO,r16,2 epoch,β=0.1) | **199/212 = 93.9%** | 0.201 | 0.384 | 0.611 | **0.357** | 512 | 56 min(2 卡)|
+| **DPO_exec_v2**(在 DPO_v1 上再做一轮:587 对新偏好对) | **204/212 = 96.2%** | 0.200 | 0.383 | 0.607 | 0.365 | – | 21 min(1 卡)|
+| **DPO_exec_v1 best-of-4**(T=0.7 采样 4 次 + 执行筛选) | pass@4 **210/212 = 99.1%**(单样本 86–93%,均值 89.6%) | – | – | – | **0.524** | – | 推理 4×0.9 min |
+| LoRA-v1 @ step 100 / 126(≈1.6 epoch) | **168/212 = 79.2%** | 0.198 | 0.385 | – | **0.301** | 705 | 24 min |
+
+初步解读:
+* 5k 条 in-domain 样本的 LoRA 就把可执行率从 0% 拉到 75%,几何指标也显著好于 Coder-7B 基线 —— **"能不能 train 起来"答案是肯定的,而且收益很大**。
+* 加入测试同名 factory(Infinigen 风格长脚本)后,模型明显更"敢写"(平均输出 774→1360 token),成功题的几何相似度更高([email protected] 0.38→0.41),但可执行率反而降到 62%(长脚本更容易在某一行撞 API 错误)。v2b(16k cutoff,纳入 347 条测试同名 factory)把这个趋势推到极致:平均输出 2580 token,成功题的 [email protected] 最高(0.414),但可执行率只有 53.8%、F(all) 最低(0.22)→ **长脚本数据在"会写复杂物体"和"写出来能跑"之间是 trade-off,当前量级下不划算;要吃下长脚本必须配执行验证 + 更大数据量(或 RL/拒绝采样)**。
+* **全参 vs LoRA**:全参三次(2 卡 offload 1 epoch、3 卡 2 epoch、4 卡 2 epoch)= 71.2 / 71.2 / 72.6%(F(all) 0.27 / 0.26 / 0.27),train loss 0.165/0.206 明显低于 LoRA 的 0.26 但测试不更好 —— 在 5k 样本量级上**全参没有额外收益,且 2 卡 offload 方案慢 10 倍(65 s/step vs LoRA 15 s/step 单卡)**。全参只有在数据量上到几十万样本/上亿 token 时才值得(届时至少 4 卡 ZeRO-3 无 offload)。
+* **数据效率很高**:v3_bio 只训 50 步(≈3M token,≈1 epoch,12 min 单卡)就到 71.7%,与 100 步的 73.6% 接近 → "学会 bpy API"在最初几百步内完成;之后的提升要靠数据质量/难度而不是更多 epoch。
+* **模型大小**:Qwen3.5-4B 同配置 LoRA 得 69.8% / F(all) 0.262(9B:75–79% / 0.28–0.30)→ 4B 拿到大部分收益,训练只快 1.2×(瓶颈不在参数量),推理显存减半;预算紧时 4B 可用,但 9B 仍稳定更好。
+* **随机性(重要)**:同一 v1 配置三次(seed/卡数不同)得 **75.0 / 78.3 / 82.1%**,F(all) 0.285 / 0.302 / 0.331 → 单次实验的噪声约 **±4 个点**。因此 v4(74.1)/v5(67.5)/v8(76.4)/v9(75.9)/v7(69.3)/4B(69.8)/全参(71.2×2)/Qwen3-8B(76.4) 这些与 v1 族的差异大多在噪声带内;真正明显的是 v6(65.6,只用外观描述)、v2/v2b(61.8/53.8,长脚本)和 best-of-n(91%)。要下结论的消融应至少跑 3 个 seed。
+* 学习率 1e-4 与 2e-4(v9:75.9% / 0.286)无差别;LoRA 对 lr 不敏感。
+* LoRA rank 16(v8)与 rank 64(v1)结果相同(76% vs 75%,F(all) 0.284 vs 0.285):这个任务/数据量下低秩就够,r16 更省显存(trainable 43M vs 173M)。
+* 4 epoch(v7)比 2 epoch(v1)略差(69% vs 75%,F(all) 0.26 vs 0.29):小数据上 2 epoch 足够,再多开始过拟合;train loss 0.20 vs 0.26 但测试没涨。
+* **可执行率与输出长度强相关**(所有 LoRA 变体一致):输出 <500 token 的题 85–92% 能跑,500–1000 token ≈70%,1000–2000 token ≈25%,≥2000 token **≈0%**(v2_indist 有 43 题输出 >2000 token,全部失败,这就是它 exec 率低的原因)。目前模型只会可靠地写短平快的 primitive 拼装脚本;复杂物体需要长脚本时几乎必错 → 下一步的数据/训练重点是**长脚本的正确性**(长样本 + 执行过滤 + 可能的拒绝采样/RL)。
+* **执行验证的自举数据(拒绝采样)**:用 LoRA-v1 采样(T=0.8,每 prompt 2 次)+ Blender 执行过滤,通过率很高:对 5193 条训练 prompt 9187/10386 = 88.5% 通过,其中 **8291 条是与训练代码不同的新解**(只有 38 条逐字复述训练脚本 → 模型并没有背题);对 9876 条**新 prompt**(articraft/deepcad 的物体描述改写成 Blender 任务)16849/19752 = 85.3% 通过,去重后得到 **15704 条新样本(覆盖 9069 个新 prompt)**。整个采样+过滤在 4 卡 + CPU 上约 50 min,即 **一小时能造出 1.5 万条经执行验证的新训练对** —— 这是扩数据最便宜的路径。用它们回灌训练的结果:v11b(+8.3k 同 prompt 新解)80.7% / F(all) 0.318,v13(+15.7k 新 prompt 样本,数据扩 4 倍)76.9% / 0.308 —— **都落在 v1 族的噪声带内(75–82%),没有带来额外提升**:自举样本仍是模型自己会写的"短平快"脚本,多了 3–4 倍同分布数据并不能教会它新的(长/复杂)能力,它们的价值在于 best-of-n 推理和作为执行反馈 RL 的数据,而非 SFT 增量(注:第一版构建脚本里 bash heredoc 把 ```python 围栏里的反引号当命令替换,导致"只剩 1 条新样本"的假象,已改成独立 python 脚本重建)。
+* **模型对比**:Qwen3-8B(2025)同配置 LoRA 76.4% / F(all) 0.277,可执行率与 Qwen3.5-9B 族持平、几何略差([email protected] 0.364 vs 0.38–0.41)→ 新旧模型 SFT 后差距不大,Qwen3.5 的优势主要在 zero-shot 输出更长更完整;工程上 Qwen3-8B 零摩擦(无需 fla/triton 特殊处理)。
+* **执行反馈 DPO(最大的单模型提升)**:用两轮采样里"同一 prompt 下执行通过 vs 报错"的 2590 对偏好对,在 LoRA-v1 合并模型上再训一个 r16 LoRA(DPO,β=0.1,lr 5e-6,2 epoch,偏好准确率从 ~0.6 升到 0.84),**greedy 单样本可执行率 75–82% → 93.9%**,[email protected](all) 0.285 → 0.357(成功题的几何质量不变,提升全部来自"不再写错"),输出也更短(512 token)。这是最便宜有效的路线:采样 + Blender 过滤 ≈ 1 小时造偏好对,DPO 训练 1 小时。叠加 best-of-4 后 **pass@4 = 99.1%、[email protected](all) 0.524**。迭代第二轮:DPO 模型在 9.9k 新 prompt 上的采样通过率已达 96.4%(v1 为 85%),只能凑出 587 对新偏好对(模型很少再出错)→ 第二轮结果 96.2% / F(all) 0.365(vs 第一轮 93.9% / 0.357),仍有小幅提升但接近饱和(模型在新 prompt 上已 96% 可执行,新的偏好对越来越少)。
+* **best-of-n 实测**:对 LoRA-v1 用 T=0.7 采样 4 次(每次 ~1.4 min 生成 + ~8 min Blender 执行),单样本可执行率 64–76%(均值 70.6%,略低于 greedy 的 75%),但**任一样本可执行 = 91.0%(4 样本)/ 94.3%(7 样本),按执行结果挑最好的 [email protected](all) 从 0.285 → 0.462(n=4)→ 0.528(n=7)**,曲线还没饱和。执行验证是便宜且有效的外环 → 推理侧上 best-of-n,训练侧用它做拒绝采样/自举数据(用模型自己生成→Blender 过滤→回灌训练)是最直接的下一步。
+* 不同变体成功的题集合高度重叠(任意两两 ∩ ≈110–145 题,12 个 LoRA 变体都能解的有 49 题),但**并集达到 204/212**,只有 8 题(BlenderRock、Boulder、几种 Coral、Carnivore 等噪声/生长类几何)从未成功 → 单次 greedy 的 75% 远不是模型上限:**多采样 + Blender 执行筛选(best-of-n / 拒绝采样)** 能把可执行率推到 90%+,也正是做执行反馈 RL 的好起点。参考代码长度并不能预测失败(never-solved 题的参考代码中位数 279 行 vs 全体 363 行)。
+* 微调后的失败主因变成长尾 API 错误(不存在的 operator、`CloudsTexture.size`、`Material.transmission_weight`、`mathutils.noise.noise_detail` 等),以及少量空输出/超时 —— 这些是"更多、更新版本(Blender 5.x)的正确样本"能继续压下去的。
+
+**基线观察**
+* Qwen2.5-Coder-7B-Instruct:18/212(8.5%)可执行;失败主因是幻觉 API(`bmesh.ops.create`、`bm.utils.*`)、漏 import(`cos`/`Matrix`)、bmesh 索引表未 `ensure_lookup_table()`、错误 kwargs。平均输出 1040 token,偏短。
+* Qwen3.5-9B(non-thinking):**0/212**。165/212 的脚本第一行就是不存在的 `bpy.ops.scene.clear()`(系统性幻觉),其余是 `from math import random`、`Scene.unit_scale` 等。平均输出 3520 token(比 Coder-7B 长得多、结构更完整),但一个幻觉调用让全部脚本报错 → 只看"能否执行"会严重低估它;我另做了一个 lenient 版本(把 `bpy.ops.scene.clear()` 替换成正确的清场代码后重跑)作为参考。
+* Qwen3.5-9B lenient(仅修正 `scene.clear` 那一行)仍只有 4/212:后面接着撞到 Blender 5.0 的 API 变化 —— Principled BSDF 的 `inputs["Subsurface"]`/`"Specular"` 已改名(KeyError 19)、颜色给 3 元组而不是 RGBA 4 元组(15)、bmesh `ensure_lookup_table()`(7)、`bm.ops`/`bmesh.ops.xxx` 不存在(12)、`Scene.unit_scale`(6)、`from math import random`(5)……即典型的"API 版本漂移 + 幻觉"长尾。
+* 结论:**bpy 这种小众 API 的"最后一公里"正确性(特别是 Blender 5.x 的 API 变化)正是 SFT 最容易补的**,few-k 样本就应能消除此类系统性错误;评测必须真的在目标 Blender 版本里执行,不能只看代码"像不像"。
+
+## 6. 结论与建议(large-scale finetune)
+### 6.1 训练侧(H100 实测)
+* **LoRA(r64,全部 linear,视觉塔冻结)@ cutoff 8192 + packing + FA2 + fla + grad-ckpt**:单卡 ~14.8 s / step(8 条 8192-token 打包序列 = 65.5k token)→ **≈4.4k token/s/GPU**,显存 ~60 GB(bs1/卡)。3.8M token 的一个 epoch ≈ 15 min/卡。首步要 JIT 编译 fla Triton 内核(~4–5 min,之后有 cache)。
+* **多卡 DDP 扩展(LoRA)**:同一 v1 配置 2×H100 DDP(每卡 bs1、grad_accum 4,等效 batch 不变)训练 17.5 min vs 单卡 30 min → **1.73× 加速**(0.96 vs 0.56 samples/s;首步编译与通信占去剩余部分);4 卡预计 ~3.3×。
+* **全参 + ZeRO-3,3×H100 无 offload(GPU 0,1,3)**:显存 ~60 GB/卡,**干净步速 ≈8 s/step(9×8192=74k token)≈ 9k token/s**(≈3k token/s/卡,约为 LoRA 的 0.7×),2 epoch ≈ 15 min;但前 45 步被同机的 Blender 执行(80 个 CPU 进程)拖到 50–60 s/step → **ZeRO-3 对 CPU 争用很敏感,训练时不要在同一台机器上跑大规模 Blender 执行**。**4×H100 ZeRO-3 无 offload(GPU 0–3,无其它负载):≈4.7 s/step(8×8192=65.5k token)≈ 14k token/s(3.5k/卡),126 步(2 epoch,3.8M token)≈ 10 min** —— 这是全参的"正常"速度:1 亿 token/epoch ≈ 2 h(4 卡)。
+* **全参 + ZeRO-3(2×H100 + CPU optimizer offload,GPU 2 被占时的退路)**:trainable 8.95B(视觉塔冻结),显存 **≈47 GB/卡**(bs1×8192,grad-ckpt),**≈65 s/step(8 条 8192-token 序列)≈1.0k token/s 总吞吐** —— 比 LoRA 慢一个数量级,瓶颈是 CPU 上 9B 参数的 Adam 更新 + 2 卡 ZeRO-3 的参数 all-gather。结论:**全参微调 9B 至少要 4 卡无 offload**(估计 4 卡 ZeRO-3 ≈ 3–4k tok/s),2 卡 offload 只适合验证可行性。经验公式:9.6B 参数 bf16 权重 + fp32 master + Adam ≈ 16 B/param ≈ 154 GB 优化器/参数状态 → ZeRO-3 分到 3 卡 ≈ 51 GB/卡 + 激活;**4 卡即可无 offload 全参微调 9B(≈40 GB/卡 状态)**;若只有 2 卡需要 `ds_z3_offload.json`(优化器 offload 到 CPU,2 TB 内存够用)。
+* 大规模数据(例如全量 3DCodeVerse ≈ 330M code token,加上 prompt/system ≈ 400M token/epoch)估算:LoRA 4 卡 ≈ 17.6k tok/s → ~6.3 h/epoch;全参 4 卡(约为 LoRA 的 0.6–0.7×吞吐)≈ 9–10 h/epoch。8 卡对半。
+* 关键工程要点:① 必须 packing(样本长度方差大,p50 600 vs p99 5k token),LLaMA-Factory 的 `neat_packing` 能避免跨样本注意力;② 长脚本(Infinigen 风格 10–20k token)需要 cutoff ≥16k–32k,建议按长度分桶或给长样本单独一轮;③ 用 `tokenized_path` 把 tokenize/packing 只做一次;④ Qwen3.5 需 `triton==3.7.1`(Hopper bug)、`flash-linear-attention` + `causal-conv1d`,否则线性注意力层退回 torch 实现会慢 3–5×;⑤ 评测生成用 vLLM(见 6.3)。
+
+### 6.2 数据侧(基于消融)
+1. **格式**:3DCodeVerse 的 `metadata.parquet`(code + captions)就是现成的 SFT 原料,不需要 tar;统一成 `{system, user: caption, assistant: ```python code```}` 即可喂 LLaMA-Factory(sharegpt json)。`captions.instruction` 与 3DCodeBench 的 `prompt_instruction` 风格一致,是首选 user prompt;**prompt 字段消融**:混用 `instruction`/`detailed`/原始 name(v1,75.0%)> 只用 `instruction`(v5,67.5%)> 只用 `detailed` 外观描述(v6,65.6%,F(all) 0.24)。**交叉测试**(把 3DCodeBench 的 `prompt_description.txt` 外观描述当 prompt):v1 75.9% / F(all) 0.294(与指令式下一致)、v6 68.9% / 0.262(在它自己的风格上也只是略好)、base 0% → 混合 prompt 风格训练的模型对测试提问风格最鲁棒;只用单一风格训练在任何风格上都更差。
+2. **量与质**:只用 bioinspired3d 的 5k 条短脚本(v3)就拿到 73.6%,和加了 distill/factory 的 v1(75%)几乎一样 → 在这个量级上"把 bpy API 用对"是主要收益来源,几千条干净可执行的样本就够了;再往上走要靠**更长、更复杂、且在目标 Blender 版本里验证过能跑的脚本**(Infinigen/distill 风格),否则模型只会写短平快的 primitive 拼装。
+3. **长脚本**:要教 Infinigen 风格(p50 9k token)必须把 cutoff 提到 16k–32k;8k cutoff 会把 60% 长样本丢掉。加入这类数据后模型输出更长、几何更像([email protected] 0.38→0.41),但可执行率下降 → 大规模训练时长脚本数据一定要配合执行验证/过滤(只收 Blender 5.x 跑通的)。
+4. **脏数据**:thingiverse-5-0 两类脚本都依赖外部资产(mesh json / CSG json),不能直接当训练目标;parametric 版要剥 shim。建议 3dcode toolkit 导出时提供 "self-contained code only" 标志,并把每条样本的 `exec.blender_5.x.status=ok` 作为训练入选条件。
+5. **解码与样板**:在样板/注释重复度高的子集(thingiverse 的 Customizer 参数块、shadertoy 的 URL/"Fork of" 头注释)上单独 LoRA 后,greedy 解码会进入复读循环(OpenSCAD 0/50、GLSL 138/200 顶上限),同一模型换 T=0.7 采样后分别 66% / 42% 可执行 → 训练目标里要**剥离无信息样板**(参数块折叠、去掉 URL/署名注释),评测与推理默认用采样或 repetition penalty,并监控"未在上限内结束"的比例作为过拟合/复读的早期信号。
+6. **去重/泄漏**:bioinspired3d 22% 代码重复;factories 与 3DCodeBench 同源,训练必须按 factory 名剔除;deepcad/shadertoy/articraft 与 Blender 目标是不同方言(v4 正在测跨方言混训的影响)。
+
+### 6.3 总结:大规模 finetune 的推荐配置
+* **模型**:Qwen3.5-9B(non-thinking 模板);如果后面要更大,同家族 Qwen3.5-27B/35B-A3B 在 LLaMA-Factory 里同样是 `qwen3_5` 模板,流程不变。
+* **框架**:LLaMA-Factory(env `lf`):yaml 配置 + `FORCE_TORCHRUN=1 llamafactory-cli train`;数据先 `tokenized_path` 预处理;`packing+neat_packing`,`flash_attn: fa2`,`cutoff_len` 16k–32k(长脚本),`per_device_eval_batch_size: 1`。
+* **方法**:先 LoRA(r16–64,lr 1e-4,1–2 epoch)迭代数据;数据定型后再全参(lr 1e-5,ZeRO-3,≥4 卡);**SFT 之后一定接一轮执行反馈 DPO**(采样 → 执行器过滤 → (通过, 报错) 对 → DPO r16 β0.1 lr 5e-6 2 epoch),这是本次最大的单步提升(75–82% → 94–96%),可迭代 2 轮;推理侧用 best-of-n + 执行筛选(pass@4 99%)。多语言一起训(md_mixed),评测/推理用采样解码。LoRA 单卡 ≈4.4k tok/s;4 卡 ≈17k tok/s → 1 亿 token 一个 epoch ≈ 1.6 h。
+* **数据**:以 `captions.instruction`(混合 `detailed`/name)为 prompt;只收在目标 Blender 版本执行成功的样本;剔除与 benchmark 同名 factory;去重;长样本单独分桶;跨方言(CadQuery/GLSL)数据对 Blender 指标中性,可做多任务但不会白送提升。
+* **评测**:vLLM 生成(1–2 min/212 题)+ Blender 无头执行(~10 min)+ 几何指标,整套 `eval/run_eval_vllm.sh` 一条命令;每个实验 30–60 min 内出数,适合做大量消融。
+* **下一步最有价值的方向**:(1) 长脚本正确性(执行过滤的长样本、拒绝采样/RL with 执行反馈);(2) 把 3DCodeVerse 其它来源(deepcad 128k、articraft、shadertoy)做成多方言联合训练,看迁移;(3) 用 VLM 判分补充几何指标(现在只看形状,不看材质/语义)。
+
+### 6.4 推理/评测侧
+* HF `generate`(bs 48,greedy)≈ 250–300 tok/s/卡,212 题 × ~3.5k token ≈ 45 min/卡;GPU 利用率 30–50% 属正常(带宽受限)。
+* **vLLM 0.27.1**(`vllm-0.27.1+cu129` wheel + torch 2.13.0+cu129,靠 CUDA minor-version compatibility 跑在 12.8 驱动上,env `vllm`):**212 条 prompt 并发 ≈ 3.2k tok/s/卡,1.5 min 生成完**(HF generate 要 45 min,快 30 倍)。推理/评测/数据蒸馏一律用 vLLM。Qwen3.5-9B 在 vLLM 里要 `limit_mm_per_prompt={"image":0,"video":0}` 关掉视觉输入,`chat_template_kwargs={"enable_thinking": False}` 关 thinking。`eval/run_eval_vllm.sh` 已接入实验流水线。
+* Blender 执行:212 题 32 进程并行 ≈ 10 min(CPU),单题超时 300 s。
+
+
+## 7. 多语言(多方言)数据可训性与迁移实验
+### 7.1 全部 3DCodeVerse 子集 → SFT 样本的可训性检查(`scripts/build_multidialect.py`)
+统一格式:system(方言专用)+ user(`captions.instruction`,deepcad 用几何描述加"Write a CadQuery script…"前缀)+ assistant(```<lang> 围栏代码)。逐条检查空代码 / <3 行 / 控制字符乱码 / 无 prompt / 重复代码,并抽样 500 条用 Qwen3.5 tokenizer 统计长度:
+
+| 方言 | 原始 | 保留 | 剔除原因 | token p50 / p90 / p99 | >8k 占比 | 估计 token 总量 |
+|---|---|---|---|---|---|---|
+| Blender Python(bioinspired3d + 62 非测试 factory) | 6,405 | 5,161 | 重复代码 1,244 | 514 / 773 / 1,575 | 0.2% | 2.9M |
+| CadQuery(deepcad 128k + articraft 6.9k) | 135,083 | 135,058 | 重复 25 | 578 / 1,820 / 17,375(articraft 长) | 3.0% | 178M |
+| OpenSCAD(thingiverse) | 1,193 | 1,190 | 短/重复 3 | 993 / 3,829 / 12,177 | 2.8% | 2.1M |
+| GLSL(shadertoy) | 119,961 | 116,096 | 空代码 238、过短 601、乱码 66、重复 2,960 | 800 / 3,024 / 9,857 | 1.8% | 164M |
+
+结论:**四种方言都能直接转成干净的 SFT 样本,没有格式/编码上的硬伤**;需要注意的只有 (a) GLSL 有 ~0.8% 空/乱码样本要过滤,(b) articraft 与 OpenSCAD 有少量 >8k token 的长样本(cutoff 16k 可覆盖 98%+),(c) deepcad 的 caption 不是指令(是几何描述),需加任务前缀,(d) shadertoy 是多文件(有 shader_json),这里只用了主 image pass 代码。LLaMA-Factory 对这些样本的 tokenize/packing 均无报错(见各 md_* 运行)。
+
+**three.js(补充)**:`threejs_distill` 340 个单文件 HTML(fable_5 100 + opus_4_8 240,captions 有 `user_instruction`/`instruction`/`style`),代码较长(token p50 4.9k / p90 11.6k / max 23.9k → cutoff 需 16k+);`threejs_repos` 是 25 个多文件真实工程(只有 docs 索引,没有逐样本 code 列)。转换后 train 280 / val 20 / test 40,格式上无问题。执行器:Playwright 无头 Chromium(SwiftShader 软件 WebGL)加载 HTML、收集 JS 报错 / 模块加载失败、检查 `<canvas>` 是否真的画了东西(踩坑:页面内 `gl.readPixels` 在合成后读到的是被清空的 drawing buffer,全部误报 blank;改为对 Playwright 截图做像素统计——量化后 >6 种颜色且 std>4 判为非空)。修好后 **40/40 测试参考代码全部 OK**(三角形/网格渲染正常,平均 ~16 s/样本,6 并发)—— 模型结果见下。
+
+| three.js(40 题 held-out,贪心,max_new_tokens 12k) | 执行 OK | 生成完成(未截断) | 输出长度 p50/p90 tok |
+|---|---|---|---|
+| Qwen3.5-9B 基座(零样本) | **31/40 = 77.5%** | 39/40 | 2.8k / 3.6k |
+| + LoRA md_threejs(280 样本,cutoff 16k,3 ep,r16,train loss 0.49,20 min/1 GPU) | 12/40 = 30% | 23/40 | 5.3k / 12.1k |
+| 同上,T=0.7 采样 | 20/40 = 50% | 36/40 | 4.0k / 12.1k |
+| Qwen3.5-9B 基座,T=0.7 采样 | 27/40 = 67.5% | 40/40 | 2.9k / 3.7k |
+
+结论:**three.js 这条链路是 work 的**(数据能转成 SFT 格式、16k cutoff 训练正常、Playwright 执行器对 40/40 参考代码全部 OK),但 **用 280 条 Claude 蒸馏样本做 SFT 反而把执行率从 77.5% 打到 30%**,原因很清楚:(a) 蒸馏代码很长(训练集 p50 4.9k / p90 11.6k tok),LoRA 学会了这种"写 400 行完整页面"的风格,17/40 在 12k 生成上限处被截断(截断的 11 条连 `<canvas>` 都还没出现,6 条页面半残截图失败);(b) 在 23 条写完的里面也只有 12 条能跑(52%,仍低于基座),错误全是 JS 细节(`Cannot access 'g' before initialization`、`THREE.Map is not a constructor`、重复声明、空引用),即模仿"长而复杂"的代码会放大出错面;而基座本来就很熟 three.js,2–4k tok 的简洁页面成功率高。对 3DCodeVerse 的启示:three.js 子集**不适合直接当 SFT 目标**(样本少、又长又复杂、基座已经会),更适合当执行反馈(DPO)的 prompt 来源,或只取短样本/压 cutoff+生成预算 ≥ 24k 再用;**32k 生成预算复测**(max_model_len 32768 / max_new_tokens 30000):仍是 12/40 = 30%,而且 **15/40 到 30k token 还没停**——看了尾巴全是复读循环(一条 3718 行里 3630 行是 `key.shadow.radius = 4;`,另一条无限生成 `iridescentLightGreen9 / iridescentDarkBlue9 …` 材质变体),即不是"代码太长被截断",而是和 OpenSCAD/GLSL 单语言 LoRA 一样的 **greedy 复读/不终止** 病灶(小数据 + 长样本 + 重复结构特别容易诱发);写完的 25 条里 12 条能跑(48%)。**T=0.7 采样对照**:LoRA 20/40 = 50%(36/40 能写完,复读大减;写完的里 56% 能跑),基座 T=0.7 为 27/40 = 67.5%(基座 greedy 77.5%)→ 采样把 LoRA 从 30% 救到 50%,但**依然低于基座**,结论不变:这 280 条长蒸馏样本做 SFT 对 three.js 是负收益;three.js 要进训练集得靠 (a) 更多、更短、去重复结构的样本,(b) 混进多语言大集而不是单独训,(c) 执行反馈 DPO 治复读。
+
+
+### 7.2 各方言执行器与留出测试集
+| 方言 | 执行器 | 指标 | 测试集 |
+|---|---|---|---|
+| Blender | Blender 5.0.1 headless(已有) | 可执行率 + 与参考网格的 Chamfer/F | 103 条留出(bioinspired3d 风格)+ 3DCodeBench 212 |
+| CadQuery | `cadquery` 2.8.0(pip,OCP) | 可执行率(得到 `result` 实体并导出 STL)+ 与参考 STL 的 F-score | 200 条留出 |
+| OpenSCAD | OpenSCAD 2026.08.19 AppImage 无头导出 STL | 同上 | 50 条留出 |
+| GLSL | glslangValidator 16.5(Shadertoy 前导 + mainImage 包装) | 编译通过率(无渲染器,暂不比图) | 200 条留出 |
+
+### 7.3 训练矩阵(LoRA r16,cutoff 8k,2 epoch)
+md_blender(4.9k)/ md_cadquery(9.8k)/ md_openscad(1.1k)/ md_glsl(9.8k)/ **md_mixed(四者合并 25.6k)**,每个模型都在 4 个方言测试集 + 3DCodeBench 上评测 → 单语言 vs 混合、跨语言迁移矩阵。
+
+**参考代码自检**(执行器可信度):CadQuery 199/200、OpenSCAD 50/50、GLSL 178/200 编译通过、Blender 留出 96/103 —— 执行器对各自语言的真实代码基本全能跑,失败的 GLSL 参考是用到了 Shadertoy 特有的 cubemap/keyboard 通道等前导没覆盖的特性。
+
+| 模型 | CadQuery exec / [email protected](all) | OpenSCAD exec / [email protected](all) | GLSL 编译 | Blender 留出 exec / F | 3DCodeBench exec |
+|---|---|---|---|---|---|
+| Qwen3.5-9B base | 20/200 = 10% / 0.023 | 21/50 = 42% / 0.196 | 78/200 = 39% | 1/103 | 0% |
+| LoRA-v1(只训 Blender) | 58/200 = 29% / 0.081 | 14/50 = 28% / 0.109 | 57/200 = 28.5% | 96/103 = 93% / 0.725 | 75–82% |
+| md_cadquery(只训 CadQuery 9.8k) | **146/200 = 73% / 0.223** | 7/50 = 14% / 0.064 | 6/200 = 3% | 0/103 | **0/212** |
+| md_openscad(只训 OpenSCAD 1.1k) | 53/200 = 26.5% / 0.064 | greedy:**0/50**(全部陷入参数声明复读循环、顶到 4k 上限);**T=0.7 采样 + 8k 上限:33/50 = 66% / 0.210**(基线 42% / 0.196) | 2/200 = 1% | 17/103 = 16.5% | 8/212 |
+| md_glsl(只训 GLSL 9.8k) | 14/200 = 7% / 0.020 | 2/50 = 4% | greedy:59/200 = 29.5%(4k)/ 56/200 = 28%(8k)——138/200 顶到上限、在 shadertoy 头部注释(URL/"Fork of …")上复读;**写完的 62 条里 56 条能编译(90%)**;**T=0.7 采样 + 8k:84/200 = 42%**(基线 39%),仍有 63 条写不到 `mainImage` | 3/103 | 0/212 |
+| md_blender(只训 Blender 4.9k,instruction prompt,r16) | 84/200 = 42% / 0.109 | 5/50 = 10% / 0.024 | 73/200 = 36.5% | 91/103 = 88% / 0.647 | 135/212 = 63.7% / 0.23 |
+| **md_mixed(四语言混训 25.6k,r16)** greedy | 157/200 = 78.5% / 0.239 | 3/50 = 6%(复读) | 37/200 = 18.5%(复读) | **94/103 = 91% / 0.708** | 142/212 = 67.0% / 0.232 |
+| **md_mixed,T=0.7 采样 + 8k** | **196/200 = 98% / 0.273** | **35/50 = 70% / 0.261** | **85/200 = 42.5%** | – | – |
+| md_mixed + 多语言执行反馈 DPO(581 对:GLSL 266 / OpenSCAD 238 / Blender 64 / CadQuery 13;cutoff 4k)greedy | 160/200 = 80% / 0.246 | 7/50 = 14% / 0.063 | 49/200 = 24.5% | 93/103 = 90% / 0.694 | 139/212 = 65.6% / 0.22 |
+| 同上,T=0.7 采样 | 196/200 = 98% / **0.282** | **36/50 = 72% / 0.266** | **94/200 = 47%** | – | – |
+| **md_big(四语言大集 101.9k 样本 / 107M tok·epoch,r64,2 epoch,4 卡 3.4 h)** greedy | **190/200 = 95% / 0.295** | 2/50 = 4%(复读) | 84/200 = 42% | 93/103 = 90.3% / 0.705 | **180/212 = 84.9% / 0.329** |
+| md_big,T=0.7 采样 + 8k(ckpt 2200,最终 ckpt 见 §8) | **195/200 = 97.5% / 0.298** | 32/50 = 64% / 0.308 | **119/200 = 59.5%** | – | 3DCodeBench pass@1 84.1% / **pass@4 98.1%(F 0.503)** |
+
+**跨语言迁移(先看 v1)**:只训 Blender-Python 的 LoRA 让 CadQuery(同为 Python、同是 CSG 建模)从 10% 升到 29%,但让 OpenSCAD 从 42% 跌到 28%、GLSL 从 39% 跌到 28.5%(20/200 连 `mainImage` 都不写了,开始往 Python 风格跑)→ **单语言 SFT 对相近语言正迁移、对远语言负迁移/遗忘**;要多语言就得混训(md_mixed 结果见下)。同分布留出测试(bioinspired3d 风格)93%、F 0.725 远高于 3DCodeBench(分布外、更难)。
+
+**OpenSCAD 子集的可训性问题**:thingiverse 的 1.1k 条 .scad 大多是 Customizer 风格、开头是几十行 `x = 5; // [1:10]` 参数块,只训它 2 epoch(loss 仍 0.88)后模型在 greedy 解码下全部陷入参数块复读、没有一条在 4096 token 内结束(基线反而 42% 能跑)→ 但换成 T=0.7 采样解码后 66% 可执行、几何 F 0.21(都高于基线)→ 模型其实学到了 OpenSCAD,只是 greedy 解码在重复样板上容易进入循环(小数据过拟合的典型症状);**结论:该子集可训但要 (a) 去掉/折叠参数块样板、(b) 评测/推理用采样或 repetition penalty、(c) 扩大数据量**(3dcodeverse 里还有 6.9k 条 parametric 版可用的 `prompt.txt → *.scad`)并与其它语言混训。
+
+**混训结论(最重要)**:四语言混训的 md_mixed 在**每一种语言上都 ≥ 对应的单语言模型**(采样解码下 CadQuery 98% vs 73%、OpenSCAD 70% vs 66%、GLSL 42.5% vs 42%、Blender 留出 91% vs 88%、3DCodeBench 67% vs 64%),而单语言模型在其它语言上几乎归零 → **3DCodeVerse 的多方言数据可以也应该合在一起训,语言之间没有互相干扰,反而有正迁移(CadQuery 被 Blender/OpenSCAD 数据带高了 25 个点)**。唯一要注意的仍是 greedy 解码在 OpenSCAD/GLSL 上的复读(混训也没解决,需要采样/重复惩罚或清理样板)。在 md_mixed 上再做一轮**多语言执行反馈 DPO**(各语言执行器造的 581 对偏好对,主要是 GLSL/OpenSCAD),greedy 下 OpenSCAD 6%→14%、GLSL 18.5%→24.5%,采样下 GLSL 42.5%→47%、OpenSCAD 70%→72%、CadQuery F 0.273→0.282,Blender 持平 —— 方向一致但幅度小于 Blender 单语言 DPO(那里有 2590 对);**DPO 的收益与偏好对数量成正比,哪种语言错得多、造的对就多、提升就大**。
+
+基线观察:Qwen3.5-9B 零样本写 OpenSCAD 最靠谱(42% 能出网格,几何也最接近),GLSL 四成能编译,CadQuery 只有 10%(API 幻觉:不存在的方法、`moveTo` 参数、`Plane` 构造),Blender 几乎全军覆没(前文的 `bpy.ops.scene.clear()` 问题)——说明各语言"API 正确性"都是 SFT 的低垂果实。
+
+## 8. 更大规模实验 md_big:训练量 ×4、四卡、更长时间("多 train 一些效果是否更明显")
+
+**设置**:把 §7 的四语言数据放大到 **101,851 样本 / 107M token·epoch**(CadQuery 40k、GLSL 40k、OpenSCAD 全部 1.1k、Blender 5.0k + 15.7k 执行验证过的自举样本;md_mixed 25.6k 的 4 倍),LoRA **r64**(md_mixed 是 r16),lr 1e-4,cutoff 8192 + packing,**4 张 H100 DDP**,有效 batch 8 × 8192 tok,**2 epoch = 3292 步 ≈ 214M token**,每 220 步存 checkpoint + 在 522 条验证集上算 eval loss。
+
+**训练侧实测**:4 卡 LoRA 稳定在 **≈17.4k tok/s(4.35k tok/s/卡)**,2 epoch 共 **3 h 25 min**(12305 s),每卡显存 54 GB;train loss 0.43、eval loss 0.44(md_mixed 同期 eval loss 更高,见 §7);机器上其他用户把 CPU 负载打到 ~300 时 GPU 利用率会从 100% 掉到 ~77%(数据侧变慢),属于共享机噪声。整条链路(建数据 → 训练 → 合并 → 3DCodeBench + 四方言评测 → 检查点扫描)全自动跑完、没有任何报错。
+
+**结果(对照 §7.3 表)**:
+
+| 指标 | md_mixed(25.6k,r16) | **md_big(101.9k,r64)** | 变化 |
+|---|---|---|---|
+| 3DCodeBench 执行率(greedy)/ [email protected] | 67.0% / 0.232 | **84.9% / 0.329** | **+18 pt** |
+| 3DCodeBench T=0.7:pass@1 / pass@4 / best-of-4 F | – | 84.1% / **98.1%** / 0.503 | v1 是 91.0% / 0.462 |
+| CadQuery 留出 200(greedy)| 78.5% / 0.239 | **95% / 0.295** | +16.5 pt |
+| CadQuery 留出(T=0.7) | 98% / 0.273 | 97.5% / 0.298 | 持平(已饱和) |
+| GLSL 留出 200(greedy / T=0.7) | 18.5% / 42.5% | **42% / 59.5%** | +23 / +17 pt |
+| OpenSCAD 留出 50(greedy / T=0.7) | 6% / 70% | 4% / 64% | 持平(只有 1.1k 条,仍靠采样救) |
+| Blender 留出 103(greedy) | 91% / 0.708 | 90.3% / 0.705 | 持平 |
+
+**结论**:
+1. **训练量确实带来明显、而且是"全面"的提升**:3DCodeBench 从 67% → 85%(超过了只训 Blender 的 v1 最好成绩 82%,且这是一个同时会四种语言的模型),CadQuery greedy 78.5% → 95%,GLSL 18.5% → 42%(greedy)/ 42.5% → 59.5%(采样)。Greedy 下的提升尤其大,说明更多数据主要是**压掉了复读/卡在头部注释这类"写不完"的失败模式**,让贪心解码也能用。
+2. **饱和的地方不会再涨**:CadQuery 采样已 98%、Blender 留出 91%,md_big 没有再动;OpenSCAD 只有 1.1k 条、在大集里占 1%,greedy 仍复读 → **数据量少的语言需要单独加权/上采样,而不是指望整体放大**。
+3. **SFT 放大 4 倍(85%)仍然不如在 5k 样本 v1 上做一轮执行反馈 DPO(93.9% → 96.2%)**,两者是正交的:md_big + 执行反馈 DPO 应是下一步(md_big 的 pass@4 已 98.1%,说明"会写对"的能力在,差的是挑出来)。
+4. 训练量曲线见下(检查点扫描)。
+
+**训练量曲线**(checkpoint → 3DCodeBench greedy 执行率 / F;同一扫描脚本,每个检查点单独合并评测;save_total_limit=6 只留下了最后 6 个检查点,更早的点被轮转删掉了——这是本次的一个疏漏,下次应 save_total_limit≥10):
+
+| checkpoint(步 / epoch) | 3DCodeBench greedy | CadQuery T=0.7 | OpenSCAD T=0.7(50 题) | GLSL T=0.7 | Blender 留出 greedy |
+|---|---|---|---|---|---|
+| 2200 / 1.34 | 182/212 = 85.8% / 0.330 | 97.5% / 0.298 | 64% / 0.308 | 59.5% | 92.2% |
+| 2420 / 1.47 | **186/212 = 87.7% / 0.341** | 98.5% / 0.273 | 62% / 0.241 | 58% | 92.2% |
+| 2640 / 1.60 | 182/212 = 85.8% / 0.335 | 96% / 0.248 | **70% / 0.305** | **67%** | 91.3% |
+| 2860 / 1.74 | 178/212 = 84.0% / 0.315 | 98% / 0.263 | 48% / 0.190 | 60.5% | 91.3% |
+| 3080 / 1.87 | 187/212 = 88.2% / 0.342(scipy 修复后重跑,原 183/0.331) | 98.5% / 0.267 | 58% / 0.242 | 64% | 92.2% |
+| 3292 / 2.00(最终) | 180/212 = 84.9% / 0.326 | 98.5% / 0.271 | 68% / 0.291 | 63.5% | 93.2%(同一权重另一次评测 90.3% → Blender 留出的噪声约 ±3 pt) |
+| 参照:md_mixed 25.6k×2 ep(≈1/4 的 token) | 67.0% / 0.232 | 98% / 0.273 | 70% / 0.261 | 42.5% | 91% |
+
+读法:从 1.34 epoch 到 2 epoch 这一段曲线**已经平了**(3DCodeBench 在 84–88% 之间抖动,±2 pt 以内是评测噪声;CadQuery/Blender 早已饱和;GLSL 在 58–67% 间波动,OpenSCAD 50 题样本太小、48–70% 都在噪声里;最后 1000 步 lr 已经很小,每个点之间的差别都在噪声内)——也就是说 md_big 的收益主要来自**前 1.3 个 epoch(数据量/多样性)**,而不是"再多跑几轮";结合 §5 的 4-epoch 过拟合结果,**推荐 1–2 epoch + 更多不同数据,而不是在同样数据上多 epoch**。第 2 epoch 的 eval loss 仍在降(0.565 → 0.44)但执行率不再涨,是"loss 降 ≠ 执行率涨"的又一个例子(§6 已有同样观察)。
+
+**md_big + 多语言执行反馈 DPO(已完成)**:用 md_big 在每种语言 1.5k 训练 prompt 上各采样 2 次(T=0.8,2 卡 ≈ 40 min)→ 各语言执行器过滤 → **1052 对** (OK, FAIL)(OpenSCAD 407 / GLSL 562 / Blender 55 / CadQuery 28;md_big 自采样通过率 CadQuery 98.6% / OpenSCAD 70.7% / GLSL 63.7% / Blender 93.4%,所以"对"天然集中在还不会的语言上)→ DPO(r16,lr 5e-6,β 0.1,cutoff 4k,2 epoch,2 卡 31 min,loss 0.58)→ 评测:
+
+| 指标 | md_big(SFT) | **md_big + 执行反馈 DPO** | 变化 |
+|---|---|---|---|
+| 3DCodeBench greedy / F | 84.9% / 0.329 | **87.3% / 0.325** | +2.4 pt(Blender 对只有 55 对) |
+| Blender 留出 greedy / F | 90.3% / 0.705 | **93.2% / 0.730** | +2.9 pt |
+| CadQuery greedy / T=0.7 | 95% / 97.5–98.5% | 94% / 98.5% | 持平(饱和) |
+| GLSL greedy / T=0.7 | 42% / 58–67% | **48% / 72.5%** | +6 / **+8–14 pt** |
+| OpenSCAD greedy / T=0.7 | 4% / 48–70% | 10% / **82% (F 0.321)** | greedy 仍复读;采样 **+12 pt 以上**,是全程最好的 OpenSCAD 结果 |
+
+结论:执行反馈 DPO 在更大的 SFT 底座上依然有效,而且**收益正好落在"对"多的语言上**(GLSL、OpenSCAD),说明 DPO 对的配比决定了它改哪里;Blender/3DCodeBench 只涨 2–3 pt 是因为 md_big 自采样 Blender 已 93% 通过、只凑出 55 对——要像 v1 那样把 3DCodeBench 推到 95%+,应该**专门为 Blender 多采样几轮(K=4–8、更多 prompt)凑到 1–2k 对**再做一轮,或按语言分轮迭代 DPO(§5 的 round-2 经验)。greedy 下 OpenSCAD 的复读仍没解掉(1.1k 条训练数据的问题,参见 §7),推理侧应统一用采样 + 重复惩罚。
+
+**§8 总结**:同样的 4 张 H100,一条 "10 万样本 SFT(3.4 h)→ 执行反馈 DPO(1.5 h 含采样)" 流水线,得到一个四种 3D 语言都能用的模型:3DCodeBench 87%(pass@4 98%)、Blender 留出 93%、CadQuery 98%、GLSL 72%、OpenSCAD 82%(后三者采样解码)。相比 8 小时前的 v1(只会 Blender、75%),**数据量(×4)和执行反馈是两个真正起作用的杠杆,epoch 数、rank、lr 都不是**。
+
+## 9. 继续 scale up:md_xl(全部可训数据,27 万样本,4 卡 1 epoch)
+
+**设置**:去掉 md_big 的每语言 4 万条上限,用全部 3DCodeVerse 可训样本:CadQuery 130.6k(deepcad + articraft 全部)、GLSL 114.2k(shadertoy 全部)、Blender 5.0k + 15.7k 自举、OpenSCAD 1.1k **上采样 ×4 = 4.4k**(针对"数据太少 → greedy 复读")、three.js 220(≤8k token 的那部分,看混训能否救它)→ **270,173 样本 / 293.4M token**(md_big 的 2.7 倍不同数据)。LoRA r64、lr 1e-4、cutoff 8k、4 卡 DDP、**1 epoch(4505 步)**——依据 §8 曲线"1.3 epoch 后就平,收益来自新数据"。每 450 步存检查点(这次留全 11 个)。建数据用新的 token 数缓存只要 1 分钟(`scripts/cache_ntok.py`,README 有说明);踩到一个新坑:GLSL 全量里 2 条文本含字面 `<image>`,qwen3_5 多模态模板把它当图片占位符直接报错 → 数据侧转义 `<image>/<video>/<audio>`(已写进 `build_md_xl.py`)。
+
+**训练侧**:4 卡 LoRA **4 h 46 min**(17168 s,≈17.1k tok/s,与 md_big 持平),train loss 0.445(1 epoch 结束;md_big 2 epoch 末 0.433),每卡 54 GB。
+
+**结果(greedy,对照 md_big / md_mixed)**:
+
+| 指标 | md_mixed 25.6k | md_big 101.9k(2 ep) | **md_xl 270k(1 ep)** |
+|---|---|---|---|
+| 3DCodeBench 执行率 / [email protected](all) | 67.0% / 0.232 | 84.9% / 0.329 | **90.1% (191/212) / 0.346** |
+| CadQuery 留出 greedy / F | 78.5% / 0.239 | 95% / 0.295 | **97.5% / 0.321** |
+| OpenSCAD 留出 greedy / F | 6%(复读) | 4%(复读) | **42% / 0.210**(×4 上采样把 greedy 复读压下去了,虽然还没到采样的 70%) |
+| GLSL 留出 greedy | 18.5% | 42% | 37.5%(噪声内;采样结果见曲线扫描) |
+| Blender 留出 greedy / F | 91% / 0.708 | 90.3% / 0.705 | **93.2% / 0.723** |
+
+→ **再放大 2.7 倍数据,3DCodeBench 继续 +5 pt 到 90%(纯 SFT 首次过 90%,超过 v1+DPO 的 93.9% 以外的所有 SFT 模型),CadQuery/Blender 小幅再涨,OpenSCAD 的 greedy 复读被上采样基本治好**;GLSL 持平。数据量这条曲线到 27 万样本还没饱和。
+
+**3DCodeBench 生成质量检查(不只看能不能跑)**:在 190 个执行成功的题上,[email protected] 均值 0.60 / 中位 0.62,**123 题 ≥0.5、82 题 ≥0.7**(主结构正确)、19 题 <0.2(跑起来但形状不对);这个分布与 v1 / md_big / DPO 几乎相同(均值 0.60–0.61)——**scale-up 提高的是"写对、跑起来"的比例,"跑起来之后像不像"这项 SFT 至今没怎么动**,要提升几何相似度得靠几何反馈(用 F/Chamfer 做 DPO 的偏好信号)而不是更多 SFT。下面是同机位渲染的对照(左生成、右 GT;按 [email protected] 取最好/中位/最差/随机各 4 题):
+
+<img src="data:image/jpeg;base64,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" alt="md_xl generated vs GT contact sheet" style="max-width:100%;border:1px solid #ccc">
+
+读图:最好的一档多是简单体(书、柜门、水滴、气球字);中位档"像但不对"(石头是多面体而非碎石、楼梯有台阶无栏杆、花盆没有植物);最差档是**零件错位/悬空**(SideTable 的桌面和腿分开飘着、筷子/铲子这类细长件比例失真);随机档里浴缸变成扁盘、龙虾/珊瑚只有"很多柱子"的大意、KitchenSpace 的柜体布局对了但细节缺失。也能看到 F 指标的局限:Balloon 题 GT 是"气球字 BALLOON"、生成是一排气球,[email protected] = 1.0(归一化后都是细长一条,10% 容差太宽松),所以**细长/扁平物体要看 [email protected]**,而 F 对"语义对不对"不敏感。
+
+**md_xl 补充评测**:
+- 3DCodeBench 采样(T=0.7):pass@1 82.8%、**pass@4 98.1%**(best-of-4 F 0.493)——和 md_big 一样(98.1%),说明 pass@4 在 98% 附近饱和、再涨只能靠 greedy/pass@1。
+- three.js(40 题):greedy 15/40 = 37.5%,T=0.7 21/40 = 52.5%(基座 77.5% / 67.5%;单独 LoRA 30% / 50%)→ **把 220 条 three.js 混进大集也救不回来**,它仍低于基座;three.js 这套蒸馏数据就是不适合 SFT。
+- **md_xl + 多语言执行反馈 DPO**(md_xl 自采样通过率 CadQuery 98.6% / OpenSCAD 84.2% / GLSL 64.5% / Blender 92.1% → 只凑出 909 对:GLSL 566 / OpenSCAD 250 / Blender 70 / CadQuery 23;DPO 26 min,loss 0.63):3DCodeBench 190/212 = 89.6% / F 0.348(与 md_xl 90.1% 持平——Blender 只有 70 对);Blender 留出 93.2% / F 0.778;CadQuery greedy 97.5% / F 0.331、T=0.7 98.5%;**OpenSCAD greedy 50%(md_xl 42%)、T=0.7 86% / F 0.337(历史最好)**;GLSL greedy 40.5%、T=0.7 63.5%。结论同 §8:执行反馈 DPO 的收益落在"对"多的语言上;底座越强、自采样越少出错,(OK, FAIL) 对就越少,**对 3DCodeBench 的边际收益在 md_xl 这个水平上已经接近 0**——这正是要换成几何反馈(§10)的原因。
+
+**md_xl 训练量曲线**(每 450 步 ≈ 0.1 epoch 一个检查点,单独合并评测;3DCodeBench 为 greedy,CadQuery/OpenSCAD/GLSL 为 T=0.7 采样,Blender 留出为 greedy;为给几何反馈 DPO 让卡,扫到 ck3150 停,终点 4505 用最终模型的数字):
+
+| 步 / epoch | 3DCodeBench / F | CadQuery | OpenSCAD(50 题) | GLSL | Blender 留出 |
+|---|---|---|---|---|---|
+| 450 / 0.10 | 159 = 75.0% / 0.290 | 98% | 68% | 44.5% | 87.4% |
+| 900 / 0.20 | 159 = 75.0% / 0.286 | 97.5% | 56% | 56.5% | 91.3% |
+| 1350 / 0.30 | 176 = 83.0% / 0.321 | 95.5% | 66% | 53.5% | 92.2% |
+| 1800 / 0.40 | 182 = 85.8% / 0.329 | 98.5% | 68% | 36.5% | 89.3% |
+| **2250 / 0.50** | **193 = 91.0% / 0.343** | 96% | 78% | 59.5% | 92.2% |
+| 2700 / 0.60 | 187 = 88.2% / 0.333 | 97.5% | **90%** | 57.5% | 93.2% |
+| 3150 / 0.70 | 186 = 87.7% / 0.337 | 99.5% | **92%** | 58.5% | 93.2% |
+| 4505 / 1.00(最终) | **191 = 90.1% / 0.346** | 97.5%(greedy) | 42%(greedy) | 37.5%(greedy) | 93.2% |
+
+读法:**0.1 epoch(2.7 万样本、29M token)就到 75%** = v1 在 5k 样本上训 2 epoch 的水平;0.5 epoch 到 91% 后进入平台(87–91% 抖动),OpenSCAD 采样从 0.6 epoch 起稳定在 90%+(×4 上采样生效)。合起来和 §8 的结论一致:**数据的多样性/覆盖面在前半个 epoch 就把大部分收益兑现,后面是噪声**。
+
+## 10. 8 小时探索:几何反馈 DPO、重新配比、全参微调("数据用光之后还能怎么涨")
+
+§9 之后的判断是:3DCodeVerse 的现成代码已经用光(md_xl),执行率的 scale 曲线在变缓、几何相似度从 v1 到 md_xl 一直没动。于是用 4 张卡并行试三个不同的杠杆(都以 md_xl 为起点或对照):
+
+### 10.1 几何反馈 DPO(A):第一次让"像不像"动起来
+**做法**:训练集里 bioinspired(Blender)/ deepcad(CadQuery)每条都有参考代码 → 执行得到参考网格;让 md_xl 对 3000 条 Blender + 1500 条 CadQuery 训练 prompt 各采样 4 / 2 次(T=0.8,共 1.5 万条,单卡 35 min),全部执行(1.2 万次 Blender 运行 ≈ 2 h CPU)并对参考网格算 [email protected];每个 prompt 里 **F 最高的当 chosen、F 最低(且差 ≥0.15、最高 ≥0.25)的当 rejected**(几何对),另外 (可执行, 报错) 也配成对 → **2357 对**(几何 2111 + 执行 246;Blender 1760 / CadQuery 597;md_xl 自采样在这些训练 prompt 上通过率 92%、F 均值 0.69)→ DPO(r16,lr 5e-6,β 0.1,cutoff 4k,2 epoch,578 步,单卡 1.5 h)。
+
+| 指标 | md_xl(SFT) | md_xl + 执行反馈 DPO | **md_xl + 几何反馈 DPO** |
+|---|---|---|---|
+| 3DCodeBench 执行率 | 90.1% | 89.6% | 88.2%(噪声内) |
+| 3DCodeBench [email protected](all / 仅执行成功) | 0.346 / 0.387 | 0.348 / – | **0.351 / 0.403** |
+| 3DCodeBench [email protected](仅执行成功) | 0.602 | – | **0.627** |
+| Blender 留出:执行率 / [email protected] / Chamfer | 93.2% / 0.776 / 0.080 | 93.2% / 0.778 / – | 93.2% / **0.800** / **0.070** |
+| CadQuery 留出:执行率 / [email protected] | 97.5% / 0.330 | 97.5% / 0.331 | 97.0% / 0.335 |
+| 3DCodeBench best-of-4(T=0.7)pass@1 / pass@4 / best-of-4 F | 82.8% / 98.1% / 0.493 | – | **85.1%** / 98.1% / **0.504** |
+
+**读法**:从 v1 到 md_xl,所有模型"执行成功题的 [email protected]"都卡在 0.382–0.389、[email protected] 卡在 0.598–0.611(§9),几何反馈 DPO 一轮就把它们推到 **0.403 / 0.627**,Blender 留出的 Chamfer 降 12%,而执行率基本不变——**这是整个项目里第一次几何相似度被撬动**,方向对了,幅度还小(一轮、2357 对、只有 Blender/CadQuery 两种语言)。它和执行反馈 DPO 是互补的:执行反馈改"能不能跑",几何反馈改"跑出来像不像",各自在自己的指标上起作用、对另一个指标中性。
+
+同一批任务(§9 那 16 题)的三列对照(左 md_xl、中 几何反馈 DPO、右 GT):
+
+<img src="data:image/jpeg;base64,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" alt="md_xl vs geometry-DPO vs GT" style="max-width:100%;border:1px solid #ccc">
+
+看图要诚实:**单任务层面起伏很大**——SideTable 从"桌面和腿分开飘"变成腿接上了([email protected] 0.02 → 0.78)、Balloon 从一排球变成更像"气球字"的环链,但 Boulder(0.62 → 0.16)、KitchenSpace(0.89 → 0.46)、SpinyLobster 反而变差,多数题基本没变。均值 +0.015–0.025 是这些涨跌的净值,说明一轮 2357 对的几何 DPO 给出的是"方向正确的弱信号",要把它变成稳定的提升需要更多对、多轮迭代、以及更能区分"零件摆对没有"的 reward(比如按零件的 F 或渲染图的 VLM 判别),而不是只看全局 F。下一步最自然的是:更多对(K=8、更多 prompt、加入 OpenSCAD)、迭代 2–3 轮、把 F 差距当权重(或直接用 GRPO 以 F 为 reward),预期还能再涨。
+
+### 10.2 重新配比 md_bal(B):Blender 份额 ×2 没用,上采样 ≠ 新数据
+**做法**:总 token 和 md_big 相同(110M),但 Blender 占比从 15% 提到 34%(bioinspired ×4 = 20.2k、自举 ×2 = 31.4k、CadQuery 35k、GLSL 25k、OpenSCAD ×4、three.js 220;共 116k 样本),LoRA r64,1 epoch,3 卡 2.45 h。
+
+| 指标 | md_big(110M tok, Blender 15%, 2 ep) | **md_bal(110M tok, Blender 34%, 1 ep)** | md_xl(293M tok 全量) |
+|---|---|---|---|
+| 3DCodeBench greedy / F | 84.9% / 0.329 | **82.5% / 0.331** | 90.1% / 0.346 |
+| Blender 留出 greedy / F | 90.3% / 0.705 | 89.3% / **0.816** | 93.2% / 0.776 |
+| CadQuery greedy | 95% | 91.5% | 97.5% |
+| OpenSCAD greedy | 4% | **56%** | 42% |
+| GLSL greedy | 42% | 24% | 37.5% |
+
+**读法**:"3DCodeBench 受限于 Blender 份额"被否了:把 Blender 样本重复 4 倍训练,3DCodeBench 没涨(82.5% vs 84.9%,噪声内偏低),CadQuery/GLSL 因为数据被削而掉;唯一涨的是 bioinspired 自己的留出 F(0.816,同分布过拟合)和 OpenSCAD greedy(×4 上采样再次证明有效)。**md_xl 的收益来自更多"不同的"代码(哪怕是别的语言),重复已有样本换不来**——要再涨只能靠新数据(合成 + 筛选)或新信号(§10.1)。
+
+### 10.3 全参微调 md_bal(C):十万级样本上全参仍不比 LoRA 好
+同样的 md_bal 数据(116k 样本 / 110M tok)、同样 1 epoch,4 卡 ZeRO-3 全参(lr 1e-5,1694 步,2 h 31 min,≈12.2k tok/s;LoRA 同数据 3 卡 2 h 27 min)。train loss 0.334 vs LoRA 0.364——loss 更低,但:
+
+| 指标 | md_bal LoRA r64 | **md_bal 全参** |
+|---|---|---|
+| 3DCodeBench greedy / F | 82.5% / 0.331 | **83.5% / 0.327**(scipy 修复后重跑,原 82.1%/0.320) |
+| CadQuery greedy / F | 91.5% / 0.279 | 91.5% / 0.302 |
+| OpenSCAD greedy | 56% | **76%** |
+| GLSL greedy | 24% | 33% |
+| Blender 留出 exec / F | 89.3% / 0.816 | 93.2% / **0.919** |
+
+**读法**:主指标(3DCodeBench、CadQuery)全参 ≈ LoRA(83.5% vs 82.5%,1 pt 以内),**在 11.6 万样本 / 1 epoch 的量级上 LoRA r64 依然没有明显容量瓶颈**——加上 §5 在 5k 样本上的同样结论,这个项目里全参从头到尾都没有赢过 LoRA,而它要多占 ~3 倍显存、快照大 30 倍。全参更低的 loss 变成了**更狠的同分布拟合**:bioinspired 留出 F 0.919(LoRA 0.816)、OpenSCAD/GLSL 的 greedy 复读也压得更低(76% / 33%)——对"贴着训练分布"的目标有用,对分布外(3DCodeBench 的 factory 风格)没用。(评测事故记录:全参模型第一次 CadQuery 评测 vLLM 引擎初始化失败给了 0/200,重跑后 91.5%——**遇到"归零"的分数先查评测日志再下结论**。)
+
+### 10.4 这 8 小时的总结
+| 杠杆 | 结果 | 建议 |
+|---|---|---|
+| 几何反馈 DPO | **有效**(scored F 首次移动:0.387→0.403 / 0.602→0.627;Blender 留出 Chamfer −12%),单任务有涨有跌 | **值得加码**:更多对(K=8、加 OpenSCAD)、迭代 2–3 轮、零件级/渲染 VLM reward,或 F-as-reward 的 GRPO |
+| 执行反馈 DPO(在 md_xl 上再来一轮) | 3DCodeBench 不再涨(底座 90% 后自采样错误太少、只凑出 909 对) | 在弱底座上很赚(v1: +19 pt),底座 ≥90% 后停用或只按语言定向凑对 |
+| 数据配比/上采样(md_bal) | **无效**(82.5%),重复样本 ≠ 新数据;只有小语种(OpenSCAD)×4 治 greedy 复读有效 | 上采样只用于 <2k 样本的语言;别指望配比解决主 benchmark |
+| 全参 vs LoRA(116k 样本) | 全参 = LoRA(主指标),只赢在同分布拟合 | 这个数据量级继续用 LoRA r64;全参等到有 >1M 样本再试 |
+| 数据 scale(§9 的延长线) | 27 万样本后曲线已缓(0.5 epoch 就 91%) | 再涨执行率靠合成+执行过滤的新数据;再涨几何靠反馈信号,不靠 SFT |
+
+## 11. 推理与评估方法论 + 全量数据 + Qwen3.8-27B
+
+前十节的模型都是"只会吐一个代码块"的微调模型,评估很省事。要评**原生的指令/推理模型**(我们真正关心的一类,比如 Qwen3.8-27B)就不一样了:它的回答里有推理、有解释、有多个代码块,而我们只要能跑的那一段代码。这一节把评估链路补成通用的,并用它测了 27B。
+
+### 11.1 从自由文本里取出"要跑的那段代码"(`eval/extract.py`)
+旧规则是"取最长的 ``` 块",在微调模型上从没出过错,但在会说话的模型上系统性出错。新的提取器:
+
+1. **先剥离推理**:`<think>/<thinking>/<reasoning>` 成对块直接删;只有开标签(生成被截断在思考里)时,把可见部分清空、再退回到思考里挖代码(记 `from_think`)。
+2. **收集候选**:所有围栏块(含最后一个未闭合的);一个围栏都没有时,从第一行"像代码"的行(`import` / `#version` / `module` / `<!DOCTYPE` / `const` …)截到结尾。
+3. **按方言打分**选块,而不是按长度:方言标志(`import bpy` / `cq.Workplane` / `void mainImage` / `THREE.` / `module …`)+3 分/条(最多 2 条)、围栏语言标签匹配 +2、语法能过(Python 用 `ast.parse`,其它用括号/结构检查)+2;shell 命令、`pip install`、纯 JSON、粘贴的报错扣分;**同分时取更靠后的块**(模型习惯把最终答案放最后),最后才比长度。
+4. 每条生成都记录提取元数据(几个块、选了第几个、有没有推理、是否无围栏、语法是否有效、有没有方言标志),聚合成 `extract_stats.json`。
+
+**回归验证**:在已有的 5 个微调模型输出目录(212+200+200+50+40 条)上,新旧提取器结果 **100% 一致** —— 也就是说前十节的所有历史数字不受影响,新规则只在"会说话"的模型上起作用。
+
+### 11.2 这件事到底有多重要(Qwen3.5-9B,同一模型开/关思考)
+`eval/eval_all.sh` 一条命令跑 6 个套件(3DCodeBench + 五个方言留出集)。同一个 Qwen3.5-9B 基座:
+
+| 套件 | 关思考 exec | 开思考 exec | 开思考时的提取情况(无围栏 / 多块 / 最佳块不在最后) |
+|---|---|---|---|
+| 3DCodeBench (212) | 0/212 = 0% | 1/212 = 0.5% | 14 / 51 / 4 |
+| Blender 留出 (103) | 1/103 = 1% | **15/103 = 14.6%** | 1 / 38 / 1 |
+| CadQuery (200) | 20/200 = 10% | **66/200 = 33%** | 1 / 156 / 5 |
+| OpenSCAD (50) | 23/50 = 46% | **42/50 = 84%** | 0 / 30 / 0 |
+| GLSL (200) | 78/200 = 39% | 70/200 = 35% | **49** / 83 / **19** |
+| three.js (40) | 31/40 = 77.5% | **35/40 = 87.5%** | 0 / 13 / 0 |
+
+结论:
+* **推理几乎在每种语言上都显著提高可执行率**(CadQuery 10→33%、OpenSCAD 46→84%、Blender 1→14.6%、three.js 77.5→87.5%),唯独 GLSL 掉了(39→35%)——因为 200 题里有 **49 题推理把 token 预算吃光、根本没写出围栏**;给足预算或关思考才公平。
+* **多块是常态**:CadQuery 200 题里 156 题给了多个代码块,GLSL 有 19 题"最佳块不是最后一块"——按最长/按最后取都会错,必须按方言打分选。
+* 3DCodeBench 仍近乎 0:那是知识问题(幻觉 `bpy.ops.scene.clear()`),推理救不了 —— **能力缺口和格式缺口要分开看**,这正是提取元数据的价值。
+* 微调模型(§5–§10 所有模型)的 `no_fence`/`multi_block` 全是 0:**SFT 顺带把输出格式钉死了**,所以线上部署用微调模型时提取几乎无风险。
+
+### 11.3 一次加载跑完所有套件(`eval/generate_multi.py`)
+27B 用 vLLM(TP=2) 加载一次要 **3.5 分钟**,原来每个套件起一次引擎 = 6 次装载。改成一个引擎顺序跑所有套件后,27B 六套件(605 个 prompt)**总共 10 分钟**(装载 214 s + 生成 9 min)。评测脚本的其余部分(执行器、几何指标)不变。踩坑记录:**不要在 bash 脚本运行时改写它**——bash 是边读边执行的,我改 `eval_all.sh` 时正在跑的两个 9B 评测在执行阶段崩了(生成结果都在,用新加的 `eval/exec_suites.sh` 补跑了打分)。
+
+### 11.4 全量数据审计:3DCodeVerse 到底有多少能用(`scripts/build_md_max.py`)
+之前只用了 27 万样本里的一部分、且每个样本只用 1 条 caption。重新审计全部 132 个 metadata 分片 + 4 个文件树:
+
+| 来源 | 样本 | 能否用 | 说明 |
+|---|---|---|---|
+| deepcad v1..v1_4 | 128,201 | ✅ | CadQuery |
+| shadertoy(112 分片) | 119,961 | ✅ | GLSL |
+| articraft cadquery_single_tex×6 | 6,882 | ✅ | CadQuery |
+| bioinspired3d geo+tex | 6,343 | ✅ | Blender |
+| 3dcodebench factories geo+tex | 486 | ✅(去掉 212 个测试 factory) | Blender |
+| thingiverse/openscad | 1,193 | ✅ | OpenSCAD |
+| **blender_distill** | 263 | ✅ 新增 | Claude 蒸馏的 Blender 脚本 |
+| **threejs_distill** | 340 | ✅ | 之前已用 |
+| **animation2code** | 1,051 | ✅ **新增语言 `web`** | 网页动画(CSS/JS/3D),带视频渲染 |
+| **3dcodebench instances_geo/tex** | 3,904 | ❌ **测试泄漏** | 100% 是那 212 个 benchmark factory 的换种子实例 |
+| articraft urdf_geo_only/tex | 6,146 | ❌ 暂不可用 | metadata 里 `code` 是空的(URDF 在 51 GB tar 里),且我们没有 URDF 执行器 |
+| threejs_repos | 25 个工程 | ❌ | 多文件工程,不是单文件 prompt→code |
+
+**关键增量:每个样本有 3 条 caption(instruction / detailed / factory)**,之前只用了 1 条。全部展开、去重、按 ≤8192 token 过滤后得到 **`md_max`:768,596 个 prompt→code 对 / 8.91 亿 token**(CadQuery 390k、GLSL 353k、Blender 19.6k、OpenSCAD 3.3k、web 1.6k、three.js 549)——这就是"接近半个 million"的那批数据,实际比这更多。`scripts/sample_mix.py` 可以按"每种语言给多少 token、每个样本最多几条 caption"从 md_max 里采训练集(caption 增强本身也能做消融)。
+
+### 11.5 Qwen3.8-27B
+架构与 Qwen3.5 同族(`qwen3_5`,55.6 GB),我们这套栈(LLaMA-Factory / vLLM / fla+triton)不用改就能跑。推理 TP=2 单卡 27 GB。**训练:单卡 LoRA(cutoff 8192)直接 OOM**(54 GB 权重 + 激活),要用 ZeRO-3 把参数切到 4 张卡。
+
+零样本(原生输出:一段电报体推理 + 一个代码块;顺带一提它没有 9B 那个 `bpy.ops.scene.clear()` 幻觉)——结果表见下(进行中):
+
+**零样本对照(贪心,`eval/eval_all.sh`;9B 与 27B 同一套流水线、同一提取器)**
+
+| 套件 | 9B 关思考 | 9B 开思考 | **27B 关思考** | 27B 开思考(12k 预算,见下) | 参考:微调后的 9B (md_xl) |
+|---|---|---|---|---|---|
+| 3DCodeBench (212) exec | 0% | 0.5% | **61.8%** (131) | 33.5% (71,204 题截断) | 90.1% |
+| 3DCodeBench [email protected](成功题) | – | 0.336 | 0.638 | **0.738** | 0.602 |
+| Blender 留出 (103) | 1% | 14.6% | **77.7%**([email protected] 成功题 0.624) | 75.7%(F 0.687) | 93.2%(F 0.776) |
+| CadQuery (200) | 10% | 33% | **62.5%** | 35.0% | 97.5% |
+| OpenSCAD (50) | 46% | 84% | **88%** | 52%([email protected] 0.736) | 42%(贪心)/ 90%(采样) |
+| GLSL (200) | 39% | 35% | **62%** | 19.5%(42 题无围栏) | 37.5%(贪心)/ 60%(采样) |
+| three.js (40) | 77.5% | 87.5% | **82.5%** | 0%(40/40 全部截断在 12288) | 37.5% |
+| 平均输出长度(bench) | 3.8k tok | 5.3k tok | **2.6k tok** | 12.2k tok(截断) | ~1.5k tok |
+
+三条结论:
+
+1. **换更大的基座比微调 9B 更"划算"的地方在于几何质量**:27B 零样本在 3DCodeBench 上执行率只有 61.8%(微调 9B 是 90.1%),**但它写对的那些题几何更准**([email protected] 0.638 vs 0.602)——而 F 恰恰是我们从 v1 到 md_xl 用尽各种 SFT 都没能推动的指标(§9)。也就是说:**微调解决"能不能跑",模型规模解决"像不像"**,两者互补,27B+微调应该两头都要(正在训)。
+2. **27B 零样本已经在 OpenSCAD(88%) / three.js(82.5%) / Blender 留出(77.7%) 上超过或接近我们微调过的 9B**,尤其 three.js(微调 9B 只有 37.5%,§7 里我们发现小样本蒸馏 SFT 反而有害)——**这类"基座本来就会"的语言,不该用小数据 SFT 去动它**。
+3. **给 27B 开思考在同等预算下是亏的**:12k 输出预算下 3DCodeBench 212 题有 **204 题被截断**(GLSL 123/200、three.js 40/40),执行率从 61.8% 掉到 33.5%;开思考的平均输出是 12.2k tok(关思考 2.6k),成本 4–5 倍。**代码生成这类任务,27B 应该关思考跑**;要公平比较得给 32k 预算(已排了一次 bench 的 32k 复测)。
+ 不过要公平地说另一面:**在没被截断、真正写完的那些题上,开思考的几何质量更高**——3DCodeBench [email protected] 0.738(关思考 0.638)、OpenSCAD 0.736(0.708)、Blender 留出 [email protected](all) 0.500(0.472)。所以正确的用法不是"关掉思考",而是**给够预算再开思考**:预算不足时思考纯亏,预算充足时它同时买到更高的执行率和更好的几何。
+
+**提取器在这里的作用(27B 开思考时的统计)**:bench 52/212 无围栏、103/212 多围栏、**74/212 的最佳块不是最后一块**(截断导致最后一块残缺);CadQuery 133/200 多围栏。用"最长块"或"最后一块"的朴素规则,这些题会被系统性判错。关思考时 27B 输出极干净(六个套件全部 0 无围栏、0 多围栏、100% 语法有效)——**"要不要复杂的提取逻辑"取决于你怎么跑推理,而不只是取决于模型**。
+
+### 11.6 一个把所有 Blender 分数压低的环境缺陷:Blender 的 Python 里没有 scipy
+
+微调后的 27B 第一次评测只有 49.5%,比零样本的 61.8% 还低。查失败原因时发现:**105 个失败里 51 个是 `ModuleNotFoundError: No module named 'scipy'`**。这不是模型的问题——**3DCodeBench 自己的 212 个参考脚本里就有 15 个 import scipy**(`interp1d` / `BSpline` / `KDTree` / `ndimage` / `sparse.csgraph`),而我们用的 Blender 5.0.1 自带 Python 只装了 numpy。也就是说:**任何忠实模仿训练数据(Infinigen 风格)写出 `from scipy.interpolate import interp1d` 的模型,都会被我们的环境判死**。
+
+修复:`blender-5.0.1-linux-x64/5.0/python/bin/python3.11 -m pip install scipy`(1.17.1;参考脚本用的 API 在这个版本都还在,不需要降级)。然后把**所有历史上受影响的评测重跑**(生成结果都在,只重跑执行+打分,纯 CPU):
+
+| run | 受 scipy 影响的题 | 修复前 | 修复后 |
+|---|---|---|---|
+| **27B 微调版(q27b_ft)** | 51 | 105/212 = 49.5% | **122/212 = 57.6%(+8.1 pt)** |
+| md_xl ck3150 | 1 | 186 = 87.7% | **187 = 88.2%** |
+| md_bal 全参 | 5 | 174 = 82.1% | **177 = 83.5%** |
+| v7_ep4 / md_bal LoRA / md_big_T07_s2 / md_xl_T07_s2,s3 / dpo_geo_T07_s3 / full_v2,v3 / v6_detail | 各 1–7 | – | 数字不变(这些题去掉 scipy 后仍因别的错误失败) |
+
+**结论与教训**:受影响最大的恰恰是最像训练数据的模型(27B 微调版 +8.1 pt),其它模型基本没动(它们本来就不写 scipy)——所以**这个缺陷没有改变报告里任何一个既有结论**,但它是一个典型的"执行式评估要连环境一起对齐"的教训:**基准的参考实现能跑通,只是必要条件的一半;还要保证被测模型合理使用的依赖也在**。已把 scipy 写进 `docs/SETUP.md` 的环境要求。(剩下 1 例 `interp2d has been removed in SciPy 1.14` 是模型用了已删除的旧 API,属真实错误,不修。)
+
+### 11.7 27B 六语言微调的结果与诊断
+
+**训练**:`md_27b_max4k`(从 md_max 采样:Blender 全量 + OpenSCAD 全量 + three.js + web 全量 + CadQuery 25M + GLSL 25M token,每样本最多 2 条 caption,去掉 >3900 token 的 2.5%)= **75,910 对 / 56.2M token**;LoRA r64、cutoff 4096、**DDP 四卡 3 h 52 min**(train loss 0.383)。
+
+**先说训练侧的一个重要教训**:27B 用 **ZeRO-3 比纯 DDP 慢 7.8 倍**(142 s/步 vs 18 s/步 ≈ 460 tok/s vs 3.6k tok/s)。LoRA 的优化器状态本来就极小,ZeRO-3 省的那点显存换来的是每步全量参数 all-gather,非常不划算。但 27B(54 GB) 在 80 GB 卡上做 cutoff 8192 的 LoRA 会 OOM → 折中是 **DDP + cutoff 4096**(只影响 2.5% 的样本,直接剔除而不是截断)。
+
+| 套件 | 27B 零样本 | **27B 微调后** | 微调 9B (md_xl) |
+|---|---|---|---|
+| 3DCodeBench (212) | 61.8% / [email protected](ok) 0.638 | 57.6% / **0.646** | 90.1% / 0.602 |
+| Blender 留出 (103) | 77.7% / [email protected] 0.472(all) | **90.3% / 0.790(all),[email protected](ok) 0.876、Chamfer 0.054** | 93.2% / 0.723(all) |
+| CadQuery (200) | 62.5% | **81.0%** | 97.5% |
+| three.js (40) | 82.5% | **95.0%** | 37.5% |
+| OpenSCAD (50) 贪心 | 88.0% | 20.0%(复读) | 42%(复读) |
+| GLSL (200) 贪心 | 62.0% | 34.5%(复读) | 37.5%(复读) |
+
+**读法**:
+* **几何质量创了全项目新高**:Blender 留出 [email protected](成功题) **0.876**、Chamfer 0.054(此前最好是几何反馈 DPO 的 0.800 / 0.070,SFT 最好 0.776)——**"更大的基座 + 微调"确实能把 SFT 一直推不动的几何指标推上去**。three.js 95%、CadQuery 81% 也都比零样本大幅提升。
+* **OpenSCAD / GLSL 在贪心下崩了**:`no_dialect_marker` 40/50 和 122/200,平均输出 6.6k / 5.4k token(大量顶到上限)。看原始输出就是 §7 里记录过的**样板文本复读**(GLSL 反复复制 shadertoy 头部 `// Fork of "Lips" by wyatt...`,OpenSCAD 反复写 Customizer 参数块)。这类方言必须用采样解码评测(v2 的评测已经加了 T=0.7 的一轮)。
+* **3DCodeBench 57.6% 仍低于零样本 61.8%**,原因有三条,都指向数据/配置而不是模型:(1) 训练 cutoff 只有 4096,而 3DCodeBench 的 factory 代码动辄几千 token,模型学到了"短代码"风格;(2) md_max 里 Blender 只有 6.5k 个**不同**样本(× 3 条 caption = 19.6k 对),比 md_xl 用的 5k + **15.7k 条执行验证过的自举样本**少了一大块多样性;(3) 混合里 CadQuery+GLSL 占了 74% 的 token。→ **v2 实验**(Blender 占比 40%、加回全部自举样本、共 104k 对 / 77M token)正在排队验证这三条。
+
+### 11.8 用尽全部数据 ≠ 更好:caption 增强的反面教训(9B 等算力三连对照)
+
+§11.4 建好 `md_max`(76.9 万对 / 8.91 亿 token)之后,最自然的问题是:**把所有来源 + 每个样本的 3 条 caption 都用上,在同样的算力下会不会比 md_xl 更好?** 三个 9B 实验,token 预算都锁在 ~3 亿:
+
+| 实验 | 数据构成 | 3DCodeBench | Blender 留出 | CadQuery | OpenSCAD 贪心 / T=0.7 | GLSL 贪心 / T=0.7 | three.js |
+|---|---|---|---|---|---|---|---|
+| **md_xl**(§9) | 27 万对,**1 条 caption/样本**,含 15.7k 执行验证自举样本,无 web/three.js | **90.1%** | 93.2% | 97.5% | 42% / 90% | 37.5% / 60% | 37.5% |
+| **md_max_9b** | 36.4 万对,**3 条 caption/样本**,全部来源(+web/three.js),**无自举样本** | **34.4%** ⬇️ | 92.2% | 97.5% | 4% / – | 33.5% / – | 42.5% |
+| **md_v3**(对照) | 27.3 万对,**1 条 caption/样本**,全部来源 **+ 全部自举样本** | **86.8%** | 87.4% | 96.5% | 2% / **54%** | 33% / **59%** | 32.5% |
+
+**结论**:
+* **同样的 token 预算下,把一份代码用 3 条不同 caption 重复三遍是净损失**:3DCodeBench 从 90.1% 崩到 34.4%(−56 pt),而同分布的 CadQuery / Blender 留出几乎不动(97.5% / 92.2%)——**受损的正是"分布外泛化"那一项**。换回 1 条 caption 并把执行验证的自举样本加回来(md_v3),马上回到 86.8%。
+* 这和 §10.2(md_bal 上采样无效)是同一条规律的两次独立验证:**决定 3DCodeBench 的是"有多少条不同的、能跑通的代码",不是"有多少条 prompt"**。0.9 亿 token 的语料里真正有用的是代码多样性;caption 增强只能在代码不足时当作弱增强,不能拿来顶替。
+* md_v3(含 web/three.js/全部 shadertoy+deepcad)比 md_xl 低约 3 pt,说明**再往里加更多"别的语言"对 3DCodeBench 也没有额外收益**——§9 得到的 90.1% 基本就是这套数据在 9B 上的天花板。
+* 采样解码再次救回样板类语言:md_v3 的 OpenSCAD 2% → **54%**、GLSL 33% → **59%**。
+
+### 11.9 训练侧的两个新坑(27B)
+1. **`PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True` 会让 27B 的 LoRA 训练 OOM**:报错落在 `fla/ops/gated_delta_rule` 的 Triton 反向内核里——Triton 自己分配显存、不走 PyTorch 的缓存分配器,扩展段把它需要的空间挤掉了。撤掉这个环境变量后,同一份配置在 cutoff 4096 上跑得比之前还快(10.5 s/步 ≈ 6.2k tok/s)。**在 80 GB 卡上压线跑 27B 时,不要开 expandable_segments。**
+2. 27B(54 GB) 在 80 GB 卡上做 LoRA 是**压线**的:cutoff 8192 必 OOM,4096 刚好;任何额外占用(评测进程还没完全退出、别的分配器策略)都会把它推翻。调度上要"连续多次确认 GPU 空闲"再起训——我第一次就是在评测的套件间隙误判空闲,撞车 OOM 白等了 5 小时。
+
+### 11.10 27B v2:全项目最好的模型(数据配比修正后)
+
+按 §11.7 诊断的三条原因改数据(**Blender 占比 15% → 40%、加回全部 15.7k 条执行验证的自举样本、去掉 >3400 token 的样本以适配 cutoff**):`md_27b_v2s` = **103,770 对 / 74.9M token**(Blender 41.2k、CadQuery 36.2k、GLSL 23.0k、OpenSCAD 2.1k、web 1.6k、three.js 0.3k),LoRA r64、cutoff 4096、DDP 四卡 **3 h 36 min**(6.2k tok/s),train loss **0.318**(v1 0.383)。
+
+| 套件 | 27B 零样本 | 27B v1(CadQuery/GLSL 为主) | **27B v2(Blender 40% + 自举)** | 最好的 9B(md_xl) |
+|---|---|---|---|---|
+| **3DCodeBench (212)** | 61.8% / F 0.257 | 57.6% / 0.249 | **93.4% (198) / [email protected](all) 0.380** | 90.1% / 0.346 |
+| **Blender 留出 (103)** | 77.7% / F(all) 0.472 | 90.3% / 0.790 | **92.2% / 0.821,[email protected](ok) 0.890、Chamfer 0.049** | 93.2% / 0.723 |
+| CadQuery (200) | 62.5% | 81.0% | 86.5% | **97.5%** |
+| three.js (40) | 82.5% | **95.0%** | 92.5% | 37.5% |
+| OpenSCAD (50) 贪心 / T=0.7 | 88% / – | 20% / – | 34% / **80%** | 42% / 90% |
+| GLSL (200) 贪心 / T=0.7 | 62% / – | 34.5% / – | 30% / **69%** | 37.5% / 60% |
+
+**结论**:
+1. **v1 → v2 只改了数据配比,3DCodeBench 从 57.6% 跳到 93.4%(+35.8 pt)**,超过之前所有模型(9B md_xl 90.1%、9B+执行反馈 DPO 96.2% 是另一条线)。这是"代码多样性 > 一切"的第三次独立验证(前两次:§10.2 md_bal 上采样无效、§11.8 caption 增强有害)。
+2. **几何质量创新高**:Blender 留出 [email protected](成功题) **0.890**、Chamfer **0.049**(此前最好 0.876 / 0.054 是 27B v1,SFT 9B 最好 0.776 / 0.080,几何反馈 DPO 0.800 / 0.070)。**规模 + 正确的数据配比,把 SFT 一直推不动的几何指标推上去了。**
+3. **每种语言的最佳模型不一样**:3DCodeBench 和 Blender 几何看 27B v2;CadQuery 看 9B md_xl(97.5%,27B 只有 86.5%——CadQuery 在 27B 混合里被压到 36k 对);three.js 看 27B v1/v2(92–95%,9B 微调只有 37.5%);OpenSCAD/GLSL 都要用 T=0.7 采样(27B v2 80% / 69%,是这两项的最好成绩)。**多语言一起训是对的,但每种语言的份额直接决定它的上限。**
+4. 输出长度也说明问题:v2 在 3DCodeBench 上平均只写 472 token(v1 1985、零样本 2619),因为 Blender 自举数据教会了它"用简洁的 bpy 代码直接建模",而不是模仿 Infinigen 那种几千行的 factory 代码。
+
+### 11.11 27B v2 + 执行反馈 DPO:底座越强,这一步的收益越集中
+
+用微调后的 27B v2 在四种语言各 1200 条训练 prompt 上采样 2 次(T=0.8)→ 各语言执行器过滤 → **只凑出 756 对**(OpenSCAD 301 / GLSL 409 / CadQuery 26 / Blender 20——v2 自采样通过率已达 Blender 94.2%、CadQuery 98.8%、OpenSCAD 82.5%、GLSL 70.5%,"错误样本"本身变稀缺)→ DPO(r16、β 0.1、lr 5e-6、cutoff 2048、**ZeRO-3**,35 min)。
+
+| 指标 | 27B v2 | **v2 + 执行反馈 DPO** |
+|---|---|---|
+| 3DCodeBench | 93.4% / F 0.380 | 91.5% / F 0.358 |
+| Blender 留出 | 92.2% / F(all) 0.821 | **93.2% / 0.813** |
+| CadQuery | 86.5% | 84.5% |
+| **three.js** | 92.5% | **100% (40/40)** |
+| OpenSCAD(T=0.7) | 80% / F 0.321 | **82% / 0.368** |
+| GLSL(T=0.7) | 69% | **71%** |
+
+**结论**:和 9B 的规律完全一致——**DPO 的收益精确地落在"能凑出偏好对"的语言上**(OpenSCAD/GLSL 有 710 对 → 各涨 2 pt 且几何变好;three.js 涨到满分),而 Blender/CadQuery 只有 46 对,3DCodeBench 反而回落 1.9 pt(噪声内)。**在弱底座上执行反馈 DPO 是最赚的一步(9B:75%→96.2%),在已经 93% 的底座上它只是补齐短板,不再是提升主指标的手段**——要继续推 3DCodeBench,得换几何反馈或按语言定向凑对(例如只对 Blender 用 K=8 采样多凑几轮)。DPO 训练本身在 27B 上必须用 ZeRO-3:它要同时前向 chosen 和 rejected,DDP 会 OOM,而这一步 token 量很小(756 对 × 2 epoch),ZeRO-3 的慢不构成问题。
+
+### 11.12 27B v3:把 CadQuery 加回来能不能"一个模型全都要"?(不能)
+
+v2 的 CadQuery 只有 3.6 万对(86.5%),而 9B md_xl 有 13 万对(97.5%)。v3 把 CadQuery 提到 **6.5 万对**、GLSL 降到 1.8 万,Blender 保持 2.2 万 + 全部自举样本(共 107,366 对 / ~7,000 万 token;因为显存压线,rank 64→32、cutoff 4096→3072,4 卡 4 h 05 min,train loss 0.264)。
+
+| 套件 | 27B v2(Blender 40%, r64/4096) | **27B v3(CadQuery 加倍, r32/3072)** | 9B md_xl |
+|---|---|---|---|
+| 3DCodeBench | **93.4%** / F 0.380 | 91.5% / 0.363 | 90.1% / 0.346 |
+| Blender 留出 exec / [email protected](ok) / Chamfer | 92.2% / **0.890** / **0.049** | 93.2% / 0.768 / 0.081 | 93.2% / 0.776 / 0.080 |
+| CadQuery | 86.5% | 85.0% | **97.5%** |
+| three.js | 92.5% | 90.0% | 37.5% |
+| OpenSCAD T=0.7 | **80%** | 62% | 90% |
+| GLSL T=0.7 | **69%** | 65.5% | 60% |
+
+**结论**:**把 CadQuery 的样本数几乎翻倍(3.6 万 → 6.5 万对)几乎没有改变 CadQuery 的成绩(86.5% → 85.0%),其它每一项还都略降。** 说明 27B 在 CadQuery 上的瓶颈不是数据量——9B 用 13 万对能到 97.5%,而 27B 用 6.5 万对停在 85%,差距更可能来自这次为了显存把 **rank 降到 32、cutoff 降到 3072**(CadQuery 的 deepcad 代码短,但 articraft 的长样本被切掉了),以及混合里 CadQuery 仍只占 ~35% 的 token。
+
+因此 **"一个 27B 通吃所有语言"在这轮没有实现**:目前最优组合仍是 **27B v2(3DCodeBench 93.4% + 最好的几何 + three.js)+ 9B md_xl(CadQuery 97.5%)**。要单模型通吃,下一步应该是:保持 v2 的配比不动、只把 CadQuery 追加到 10 万对以上,并把 rank/cutoff 恢复到 64/4096(需要 5 张以上的卡,或用 8-bit 优化器把显存压下来)。
+
+**当前全项目最佳一览**(截至 2026-08-25):
+
+| 目标 | 最好的模型 | 成绩 |
+|---|---|---|
+| 3DCodeBench 执行率 | 9B v1 + 执行反馈 DPO ×2 | **96.2%** |
+| 3DCodeBench(单次 SFT,无 DPO) | **27B v2** | **93.4%**([email protected] 0.380) |
+| Blender 留出几何质量 | **27B v2** | **[email protected] 0.890 / Chamfer 0.049** |
+| CadQuery | 9B md_xl | **97.5%** |
+| three.js | 27B v2 + 执行反馈 DPO | **100%** |
+| OpenSCAD(采样) | 9B md_xl / 27B v2+DPO | 90% / 82% |
+| GLSL(采样) | 27B v2 + 执行反馈 DPO | **71%** |
+
+## 12. 数据清洗、图像条件生成、泄漏探针(2026-08-26 → 08-28)
+
+### 12.1 把全部 3DCodeVerse 洗成可直接训练的 LLaMA-Factory 数据
+每个源子目录旁边生成一个 `<subdir>_llamafactory/`(`train.parquet` + `dataset_info.json` + `qc.json` + README),已上传到 HF 母数据集:**203 个目录 / 109 万行**。三个实质问题都在源头修掉,而不是绕过:
+
+| 问题 | 处理 |
+|---|---|
+| `articraft/urdf_*` 的 `code` 列全空(6,146 条) | 从 22 GB tar 里恢复**全部 6,146 个 `.urdf`(零遗漏)**,成为可训练的**第七种语言**;抽检 400 条 XML 解析 100% 通过,平均 4.4 link / 3.4 joint |
+| Shadertoy 多标签页不是单文件 | 用 `shader_json.renderpass` 按 Shadertoy 真实语义合并(Common 拼在 pass 前),**68 个子目录 / 287,679 行**,其中 23,988 个程序合并了 Common |
+| `3dcodebench/instances_*` 与基准同源 | 构造出来但**逐行标记** `benchmark_factory` / `factory_family`,并给出 factory 级切分 `benchmark_split.json`(留出 50 个 family),让这批数据可用而不污染评测 |
+
+### 12.2 GLSL:用真编译器 + 真渲染器验证,而不是静态检查
+
+| 阶段 | 可编译 |
+|---|---|
+| 原始单文件 + 粗糙静态检查 | 98,518 / 116,498 = 84.6% |
+| 合并 Common 标签页后 | 88,245 / 93,434 = 94.5% |
+| **修正验证器后(见下)+ 采纳编译验证过的修复** | **285,486 / 287,679 行 = 99.2%** |
+
+4,192 个"编译失败"里**只有 588 个真的需要改代码**(去掉非常量初始化的 `const`、重命名与内置同名的函数、补全截断语句、删除合并产生的重复声明——每条改动都重新编译验证)。其余全是**我的验证器的缺陷**:用正则猜 iChannel 是 2D 还是 cubemap(Shadertoy 里这是绑定属性,不是语法属性)、前导码漏了 `iFrameRate`、去重正则误删压缩成一行的真实代码、用了 ES 3.10 而 Shadertoy 是 WebGL2 = ES 3.00。
+
+**编译 ≠ 能画出东西**,所以又加了一层:在真实 WebGL2 里渲染每个 shader、采两帧、检查画面非均匀且随时间变化。第一版渲染器把所有 shader 判成全黑——用"必然出渐变"的对照 shader 一测就发现是截图失败吞掉了有效测量。**这是本轮第三次"先怀疑检查器、再怀疑数据"救回大批数据**(前两次:Blender 缺 scipy、three.js 的 `readPixels` 读到已清空的缓冲)。
+
+### 12.3 泄漏探针:不是所有泄漏都让分数虚高
+我自己写的 `build_md_max.py` 排除规则漏了 `Factory` 后缀(`AgaveMonocotFactory` ↔ 测试任务 `AgaveMonocot_seed0`),导致 md_max 混入 462 行、9B clean 混入 164 行基准 factory 源码(0.06–0.09%)。早期的 `build_multidialect.py` 规则是对的,**md_xl 的 90.1% 及此前全部 9B 结果不受影响**。
+
+用完全相同的配置训了一对模型来实测:
+
+| | 含 164 行泄漏 | **去泄漏** |
+|---|---|---|
+| 3DCodeBench 执行率 | 84.9% | 84.9%(一模一样) |
+| [email protected](成功题) | 0.383 | **0.418** |
+| Chamfer | 0.201 | **0.190** |
+
+**污染没有抬高分数,去掉之后几何质量反而更好。** 原因是这些 factory 代码动辄几千行、超 8192 token 被截断,污染的是**输出风格**而不是答案。结论:泄漏要按"它到底教了模型什么"来判断,长而畸形的样本本身就是噪声;同时也说明**只用执行率一个指标会完全看不到这个差异**(84.9% vs 84.9%)。
+
+### 12.4 图像条件生成:第一次让模型"看图写代码"
+`qwen3_5` 模板本就注册了多模态插件(`qwen3_vl`,image token `<|image_pad|>`),数据加 `images` 列 + prompt 里放等量 `<image>` 即可(框架会严格校验个数)。建了 **7,997 个"渲染图→代码"样本**(bioinspired3d 6,343 + animation2code 1,051 + threejs_distill 340 + blender_distill 263),单视角和 4 视角两版。
+
+| 模型 | 输入 | 3DCodeBench 执行率 | [email protected](成功题) |
+|---|---|---|---|
+| 9B 纯文本(md_9b_noleak) | 文字描述 | 84.9% | 0.418 |
+| **9B 图像条件(`lf_qwen35_9b_img4`)** | **一张 GT 渲染图,无任何文字** | **69.3%**(放开预算后为 64.6%,见 §12.8) | 0.373 |
+
+只给一张图、不给任何文字描述就能让 147/212 的代码跑起来,说明这条路是通的。训练侧的信号也很明确:图像条件训练的 loss 0.180,远低于纯文本的 0.31–0.39——**图像提供的条件信息比 caption 强得多**。下一步(已排队):4 视角 vs 1 视角对照、图文联合训练的双路评测。
+
+### 12.5 两个把多模态实验"评没了"的评测器缺陷(2026-08-29 修正)
+
+这一轮没有新模型,只有对**评测器**的两处修正——但它们直接决定了 §12.4 那张表能不能信。
+
+**(1)输出上限 6144 把"没写完"记成了"写错了"。** img4 那次 212 题里 **72 题在 6144 处被截断**,其中 36 题的报错是 `'(' was never closed`——这不是模型不会写,是它还没写完就被掐断。Qwen3.5/3.8 的上下文是 262,144,而 prompt 只有 ~300 token,设 6144 毫无道理。现在全线改成 **32768**,`eval_all_run2.sh` 也已是这个值。我当时据此判断 **§12.4 的 69.3% 是一个下界**——**这个判断后来被实测推翻了,见 §12.8**。
+
+**(2)评测前的等待条件用了训练的显存标准,于是整段评测被静默跳过。** `run_mm_ablation.sh` 在跑评测前调用 `free_gpus.sh`,而它的默认门槛是 **60 GB**——那是训练一个 9B 的需求。这台机器是共用的,另一位用户常驻约 10.7 GB/卡,条件长期不成立,脚本等满 10 分钟后直接 `continue`,日志里只留下一行:
+
+```
+[exec] status counts: {} | OK rate = 0/0 = 0.0%
+[mm-img_text] EVAL DONE
+```
+
+"EVAL DONE" 和 0/0 并排出现,看起来像模型全错,实际上**根本没有生成过一个样本**。修正有三层:门槛按评测的真实需求设为 40 GB(18 GB 权重 + 32k 输出的 KV cache);跳过时把四张卡的占用打印出来;生成目录为空时**拒绝打分**,而不是把空目录算成 0%。
+
+**(3)连带修掉的一个共用 GPU 陷阱**:vLLM 的 `gpu_memory_utilization` 是**总显存**的比例,不是空闲显存的比例。写死 0.88 等于向一张别人已经占了 12 GB 的卡索要 70 GB,报错是 `Free memory on device cuda:0 (45.06/79.19 GiB) ... is less than desired GPU memory utilization`。三个生成脚本现在都从**实际空闲显存**反算这个比例并打印原因。
+
+**这三条都是同一条老规律的新实例:先怀疑测量工具,再怀疑数据**(three.js 的 readPixels、Blender 缺 scipy、GLSL 采样器类型猜测、渲染验证器的截图 bug,现在再加上"评测被静默跳过")。它们共同的特征是**失败会伪装成一个合理的低分**——0/0 显示为 0.0%,截断显示为语法错误,跳过显示为 EVAL DONE。因此新脚本里加了一条硬规则:**产出为空时必须报错,绝不打分**。
+
+### 12.6 三路对照的结论:参考图没有替代 caption 的价值,而"何时停下"是文字给的
+
+`mm_img` / `mm_img_text` / `mm_text` 用**完全相同的样本和完全相同的划分**训练,唯一区别是给不给渲染图、给不给文字。全部在 32768 输出预算下重测(`run_mm_eval_fixed.sh`):
+
+| 变体 | 输入 | 3DCodeBench 执行率 | 截断 | 平均输出长度 |
+|---|---|---|---|---|
+| `mm_text` | 只有文字指令 | **49.1%** (104/212) | 53/212 | 8,774 tok |
+| `mm_img_text` | 渲染图 + 文字指令 | 48.6% (103/212) | **35/212** | **6,024 tok** |
+| `mm_img` | 只有渲染图 | 36.8% (78/212) | **97/212** | **15,445 tok** |
+
+**(1)已有文字描述时,再加一张参考图不涨分。** 49.1% vs 48.6% 只差一题,是噪声。这和 §12.4 结尾"图像提供的条件信息比 caption 强得多"的推测相反——那个推测的依据是训练 loss(图像条件 0.180 vs 纯文本 0.31–0.39),但**更低的 loss 只说明图像让下一个 token 更好预测,不等于生成的程序更能跑**。用 loss 判断多模态是否有用是不成立的。
+
+**(2)图像唯一改变的是输出长度,而且方向取决于有没有文字。** 图+文字是三者里最简洁的(6,024 tok,只有 35 题没写完);只给图则失控(15,445 tok,97 题到 32k 还没停)。**决定"什么时候算写完"的是文字指令**:只给一张图,模型没有范围概念,就一直往上堆细节。
+
+**(3)因此把生成上限从 6144 提到 32768,对 `mm_img` 一分没涨**(36.8% → 36.8%),只是让它多啰嗦三倍。这修正了 §12.5 的说法:**上限过低确实会把"没写完"记成"写错了",但那只对 `img_text`/`text` 成立;对 `img` 而言瓶颈从来不是截断,而是不终止**——和 §7 里 three.js 单语言 LoRA 的复读病灶同源(小数据 + 长样本 + 重复结构)。提高预算没有修好分数,它**暴露了真正的病因**,这恰恰是给足预算才能看清的东西。
+
+注意绝对值:这三个模型只有 6.4k 训练样本(且以 bioinspired3d 为主),所以 49% 远低于主线文本模型的 85–90%。三者之间的比较是严格可控的(同样本同划分),但**不能用来和主线模型比高低**。
+
+**(4)27B 多模态在这台机器上跑不起来。** DDP 下每个 rank 都要放完整的 54 GB 权重(与用几张卡无关),同机另一位用户常驻 10.5 GB/卡后,cutoff 6144 和 cutoff 4096 + `image_max_pixels: 200704`(448×448)两次都 OOM。唯一的出路是 ZeRO-3,而 §11.9 实测它比 DDP 慢 7.8 倍。结论:**27B 多模态需要独占的卡,共用机器上应该只做 9B 的多模态对照。**
+
+### 12.7 噪声底来自"换配置",不是来自贪心本身
+
+跨方言评测时发现同一个模型(`mm_text`)在两次评测里给出 **49.1%** 和 **47.2%**。查证后确认两次的**模型、prompt 文件、temperature=0、`no_think`、`max_new_tokens=32768` 全部相同**,唯一区别是走了 `generate_vllm.py` 还是 `generate_multi.py`(后者开了 prefix caching,`max_model_len` 也不同)。逐题比对:
+
+| 比较项 | 结果 |
+|---|---|
+| 生成的程序逐字节相同 | **51 / 212** |
+| 程序不同 | 161 / 212(长度可差一倍,如 11,740 vs 7,929 字节) |
+| **执行状态翻转** | **29 / 211 = 13.7%**(OK→FAIL、FAIL→EMPTY、FAIL→OK 都有) |
+| 总执行率变化 | 仅 −1.9pp(翻转方向大致抵消) |
+
+原因是 vLLM 的批处理推理**不是逐位确定的**:批次组成与 prefix caching 改变浮点归约顺序,当两个 token 的 logits 接近平局时结果翻转,之后整段生成分叉。平均 9k token 的长输出把这个效应放大到"一半以上的程序都不一样"。
+
+**但"贪心不可复现"这个说法是错的**——补做的对照(`scripts/repro_probe.sh`:同一脚本、同一批参数、同一张卡,连跑三次)给出:
+
+| 对比 | 逐字节相同 |
+|---|---|
+| run1 vs run2 | 207 / 212 |
+| run1 vs run3 | 207 / 212 |
+| **run2 vs run3** | **212 / 212(完全一致)** |
+| 上表的换入口对比 | 51 / 212 |
+
+**配置完全不变时,贪心是可复现的**(run2 与 run3 逐字节一致;run1 只差 5 题,且差异集中在几千到十万字节的超长输出上,属于首次运行时 KV-cache 状态不同)。**噪声来自配置改变**:换生成脚本、开关 prefix caching、改 `max_model_len`,都会改变批次组成与浮点归约顺序,进而在 logits 接近平局处翻转 token。
+
+所以准确的表述是:**同一评测入口内的比较是可靠的;跨入口、跨时间、跨脚本版本的比较有约 ±2pp 的噪声底**,而这恰恰是"拿今天的实验和上周的数字比"最常见的情形。据此重读 §12.6 的跨方言对照(两个模型走的是同一个 `eval_all_run2.sh`,因此这一层噪声不适用,但两个模型输出长度不同仍会改变批次组成):
+
+| 套件 | `mm_img_text` | `mm_text` | 差 | 读法 |
+|---|---|---|---|---|
+| openscad | 40.0% | 28.0% | +12.0pp | 可能是真的 |
+| bench | 53.8% | 47.2% | +6.6pp | 勉强超出噪声 |
+| glsl | 27.0% | 32.5% | −5.5pp | 噪声带内 |
+| blender 留出 | 87.4% | 92.2% | −4.8pp | 噪声带内 |
+| cadquery | 21.0% | 25.5% | −4.5pp | 噪声带内 |
+| threejs | 60.0% | 57.5% | +2.5pp | 噪声 |
+
+三升三降、且多数落在噪声带内——**没有证据表明只在 Blender 上加参考图会系统性地影响其它方言**,无论正负。
+
+反过来看,§12.6 里能站住的结论恰恰是那些**幅度远超噪声**的:`img` 36.8% vs `text` 49.1% 差 12.3pp,且伴随 97/212 不终止和 15,445 的平均长度——这不是噪声。而 `img_text` 48.6% vs `text` 49.1% 差 0.5pp,本来就在说"没有差别",噪声底的存在反而加强了这个结论。
+
+**对后续实验的要求**:(a) 一组要互相比较的模型,必须用**同一个评测入口、同一批参数**跑完,不要一半用 `generate_vllm.py` 另一半用 `generate_multi.py`;(b) 跨时间引用旧数字时,小于 5pp 的差异不足以下结论,需要用同一入口重测;(c) 报告里既有的小幅度对比按这个标准重新审视。这也解释了一个长期现象——同一模型换个脚本重跑分数总有出入,此前被归因为"环境波动",实际是**批次组成改变了浮点归约顺序**。
+
+### 12.8 放开输出上限之后:更多预算没有换来更好的代码,而"给几张图"比"给多少预算"重要得多
+
+§12.5 去掉了硬编码的 6144 上限,改为从 262,144 向下探测(`scripts/eval_img_budget.sh`;这张卡上 262,144 起不来,131,072 可以)。我当时预判受截断影响的分数会**上升**。实测相反。
+
+同一个模型 `lf_qwen35_9b_img4`,同一套 212 题,只改输入图片数和输出预算:
+
+| 输入 | 预算 | 平均输出 | 未写完 | 执行率 |
+|---|---|---|---|---|
+| **1 张图** | 原始(低) | 1,501 tok | — | **69.3%** |
+| 1 张图 | 12k | 2,287 tok | — | 64.2% |
+| 1 张图 | 131k | 21,204 tok | 34/212 | 64.6% |
+| **4 张图** | 6,144 | 2,526 tok | 72/212 | 43.9% |
+| 4 张图 | 12k | 4,666 tok | — | 46.7% |
+| 4 张图 | 131k | 42,365 tok | 69/212 | 47.6% |
+
+**(1)我在 §12.5 的预判是错的。** 1 张图这一路把预算放大 87 倍(1.5k → 131k),分数从 69.3% 掉到 64.6%,平均输出涨了 14 倍。低预算不是在压低分数,它在**强迫模型收敛**。4 张图那一路确实涨了(43.9% → 47.6%),但 +3.7pp 只勉强超过 §12.7 的噪声底,而代价是平均输出从 2.5k 涨到 42k——**多出来的 40k token 几乎没有换到正确性**。
+
+去掉硬编码上限本身是对的:一个写死的低上限会把"没写完"静默记成"写错了",不测就永远不知道差别有多大。但**测出来的答案和预期相反**,这正是要测的理由。
+
+**(2)一张参考图好过四张(64.6% vs 47.6%,同一个模型)。** 这个差距 17pp,远超噪声。四张图占掉更多上下文,模型输出翻倍(42k vs 21k)、未写完的题也更多。**多视角在这里是负收益**——和"更多输入信息一定更好"的直觉相反。
+
+**(3)三节合起来看,图像条件生成的瓶颈是"何时停下",不是"看不看得懂图"。** §12.6 的 `mm_img` 平均 15,445 token、97/212 不终止;这里 4 张图放开预算后平均 42,365 token。给更多预算、更多视角,模型都只是写得更长。下一步(已排队 `run_img_decoding.sh`)用同一个模型对比贪心 / T=0.7 / presence_penalty / repetition_penalty——**如果只换解码就能涨分,说明模型会建这些物体,只是不知道何时收尾**,那么该修的是解码或训练数据的长度分布,而不是继续加数据。
+
+## 13. 一个贯穿全项目的评测缺陷:贪心解码下的失控生成(2026-08-30)
+
+### 13.1 OpenSCAD 的 6% 从来不是"模型忘了这个方言"
+
+六方言横评里,每个微调过的 9B 在 OpenSCAD 上都是 2–6%,而**没微调的基座是 46%**(no-think);three.js 同样从 77.5% 掉到 32–52%。27B 是同一模式(基座 88% → 微调后 12–36%)。我最初判断这是灾难性遗忘——**这个判断是错的**。
+
+同一份权重(`mmmix2`)只改解码:
+
+| 套件 | 贪心 | T=0.7 采样 | 贪心平均输出 | 采样平均输出 |
+|---|---|---|---|---|
+| **openscad** | **6.0%** | **44.0%** | **30,166** | 14,797 |
+| glsl | 60.0% | 66.0% | 11,201 | 5,234 |
+| cadquery | 94.5% | 98.0% | 2,237 | 925 |
+| bench | 93.9% | 91.0% | 1,266 | 496 |
+| threejs | 45.0% | 37.5% | 12,090 | 4,277 |
+
+采样把 OpenSCAD 拉回 44%,**基本等于基座的 46%**。规律很干净:**贪心下输出爆炸的方言正好是分数崩掉的方言**。OpenSCAD 平均 30,166 token,意味着几乎每道题都撞到 32k 上限。
+
+直接看一份输出就清楚了。最长的那份"OpenSCAD 程序"206 KB、576 行,其中只有 146 行不重复,主体是同一句注释重复几百遍:
+
+```
+//The case is designed to be printed at 0.4mm gap between hinges and pencil
+ holder holder holder holder holder holder holder holder holder ...
+```
+
+**它一行几何代码都没写**,从开头的注释块就掉进循环出不来。这不是"不会 OpenSCAD",是退化重复。
+
+### 13.2 两个被自己的实验证伪的假设
+
+**假设一:配比失衡。** 训练集里 CadQuery 占 53.6%、GLSL 31.3%,而 OpenSCAD 只有 0.6%、three.js 0.2%——崩掉的正好是最少的两个,看起来因果很顺。于是重新配比(OpenSCAD 提到 4.0%、Blender 提到 23.6%,总量 2.8 万条)训了一版:
+
+| 套件 | mmmix2(20.6 万条) | 重新配比(2.8 万条) |
+|---|---|---|
+| bench | **93.9%** | 46.2%(输出 1,266 → 9,771 tok) |
+| cadquery | **94.5%** | 82.0% |
+| glsl | **60.0%** | 12.0%(输出 11,201 → 28,375 tok) |
+| openscad | 6.0% | 12.0% |
+
+**失控扩散到了所有方言。** 为了给弱方言腾出比例而砍掉数据,代价是 bench 掉 47pp,而 OpenSCAD 只从 6% 挪到 12%。配比不是杠杆。
+
+**假设二:数据量不足。** 上面那个结果很容易被读成"数据越多输出越短"。横向检验否定了它:
+
+| 训练条数 | 模型 | openscad 平均输出 |
+|---|---|---|
+| 27,949 | bal_lr1e4 | 28,886 |
+| 101,851 | clean | 7,705 |
+| 205,883 | **mmmix2** | **30,166** |
+| 768,596 | md_max | 7,720 |
+
+不是单调的——20 万条的 mmmix2 反而最严重。**数据量会放大这个问题,但不解释它。** 两次都是从一两个数据点推因果,推快了。
+
+### 13.3 干预:让模型自己生产负样本,把"不停"训掉
+
+采样只是绕过,而且不普适(three.js 反而降 7.5pp、bench 降 2.9pp)。既然 OpenSCAD 全语料只有 1,104 条、加数据这条路堵死,就直接针对**可观测的行为**训练,不依赖对根因的判断:
+
+* **负样本**:模型自己贪心跑出的失控输出(未终止,或长度超过它训练里见过的任何答案)
+* **正样本**:同一 prompt 下采样得到的、能正常终止的输出
+* **只用训练集 prompt**——OpenSCAD/GLSL/three.js 三个套件是留出测试集,绝不进训练
+
+3,687 个 prompt 产出 **1,605 对**:openscad 876(1,104 个 prompt 里 **79% 失控**)、glsl 469、web 215、threejs 45;另有 2,027 个贪心本来就正常。配对质量一目了然:**负样本中位数 46,202 字符,正样本 1,210 字符,长度比 35.4 倍**。
+
+### 13.4 评测协议的改动
+
+**从此长输出方言必须同时报告贪心和采样两组数字。** 只报贪心正是让这个缺陷藏了整个项目的原因——报告前面所有 OpenSCAD/GLSL/three.js 的贪心数字都系统性低估了真实能力,包括 §11.7 和 §11.12 里 27B 的那些。这些数字不撤回(它们如实反映了"贪心解码下会发生什么"),但**不能当作模型能力的度量**。
+
+这也顺带修正了 §12.6 的一条结论:`mm_img` 36.8% vs `mm_text` 49.1% 的 12.3pp 差距是真的,但**成因是解码不是能力**——`mm_img` 换成 T=0.7 后是 48.6%,与文本路径持平(同一次解码对照里,presence_penalty 35.4%、repetition_penalty 31.1%,两种惩罚都无效,说明这不是 token 级重复,而是贪心路径陷入高概率循环)。
+
+### 13.5 27B 多模态:能训,但不值得
+
+四次多卡尝试全部 OOM,每次都差 1–4 GB,且总是死在与他人共享的那张卡上;**单卡在唯一一张全空的 H100 上跑通,占 78.2 / 79.6 GB,余量 1.4 GB**(cutoff 4096 + `image_max_pixels: 200704`)。所以 27B 多模态 LoRA 装得下一张空卡,装不下任何被别人碰过的卡。
+
+训练 2 epoch、loss 0.136、4 小时 33 分。评测 **47.2%**(100/212),而 9B 的 `mm_img_text` 是 48.6%——**参数量三倍没有带来任何收益**,与 §12.6"参考图相对 caption 没有增量"一致。
+
+**评测侧还踩了三个坑,都值得记**:(1) vLLM 的 `gpu_memory_utilization` 按总显存算比例,对 51 GB 的模型 4 GB 余量不够,KV cache 探测阶段会 OOM——余量需按模型体积缩放;(2) 真正拦住它的其实是 `max_num_seqs`:Qwen3.5/3.8 有线性注意力层,vLLM 每个并发序列要一个 Mamba cache block,27B 只剩 179 块而默认要 1024,必须显式设 `--max_num_seqs 64`;(3) 我自己的脚本用 `2>&1 | tail -1` 把报错吞掉了,导致前两个判断都建立在猜测上——**诊断信息不能进管道**。
+
+### 13.6 结果:一小时单卡 DPO,四个方言大幅提升,代价接近于零
+
+1,605 对模型自己生成的偏好样本,LoRA r16、`pref_beta` 0.1、cutoff 4096,单卡约一小时。六方言在**同一次会话内**用两种解码各测一遍:
+
+**执行率**
+
+| 模型 | bench | blender | cadquery | openscad | glsl | threejs |
+|---|---|---|---|---|---|---|
+| mmmix2 贪心 | 93.9% | 92.2% | 94.5% | 6.0% | 60.0% | 45.0% |
+| mmmix2 T=0.7 | 91.0% | – | 98.0% | 44.0% | 66.0% | 37.5% |
+| **stopDPO 贪心** | 92.5% | 92.2% | **98.5%** | **40.0%** | **79.5%** | **80.0%** |
+| **stopDPO T=0.7** | 89.6% | 91.3% | **99.5%** | **90.0%** | **87.0%** | 65.0% |
+
+**平均输出长度**(被直接训练的那个行为)
+
+| 模型 | bench | cadquery | openscad | glsl | threejs |
+|---|---|---|---|---|---|
+| mmmix2 贪心 | 1,266 | 2,237 | **30,166** | 11,201 | 12,090 |
+| stopDPO 贪心 | 632 | 982 | 17,802 | 5,540 | 2,963 |
+| stopDPO T=0.7 | 476 | 702 | **517** | 572 | 2,764 |
+
+OpenSCAD 的输出从 30,166 token 缩到 517(**58 倍**),执行率 6% → 90%。基座是 46%,所以这不是"恢复到微调前",是**接近翻倍地超过基座**。GLSL 60% → 87%,three.js 45% → 80%(基座 77.5%),CadQuery 94.5% → 99.5%。bench 掉 1.4pp、Blender 基本不动。
+
+**没有引入任何新数据。** 负样本是模型自己贪心跑出的失控输出,正样本是它自己采样写出的正常版本,prompt 全部来自训练集。这条路之所以可行,正是因为 §13.1 的诊断把问题定位成了"行为"而不是"知识"——知识补不了(OpenSCAD 全语料只有 1,104 条),行为可以。
+
+### 13.7 噪声底:实测,而不是估计
+
+同一模型、同一入口、同一批参数,**在同一次会话内**连跑三遍:
+
+| 套件 | run1 | run2 | run3 | 波动 |
+|---|---|---|---|---|
+| bench | 47.2 | 47.2 | 47.2 | **0.0pp** |
+| blender | 92.2 | 92.2 | 92.2 | **0.0pp** |
+| cadquery | 24.5 | 24.0 | 24.5 | 0.5pp |
+| glsl | 29.5 | 31.0 | 30.0 | 1.5pp |
+| openscad | 36.0 | 34.0 | 34.0 | 2.0pp |
+| threejs | 57.5 | 60.0 | 62.5 | **5.0pp** |
+
+**这否定了我在 §12.7 的估计。** 我从 3DCodeBench 推出"±2pp 的统一噪声底",实测是:同批次内多数套件 0–2pp,只有 three.js 因为只有 40 题而达到 5pp。据此重读 §12.6 的跨方言表,**多数差异是真的**——bench +6.6pp(底 0.0)、blender −4.8pp(底 0.0)、cadquery −4.0pp(底 0.5)都站得住,只有 glsl −0.5pp 和 threejs +2.5pp 属于噪声。我当时用一个没验证过的猜测把真实信号一并否掉了,方向反了。
+
+还有一层:同一模型同一入口的 `mm_text`,**跨批次**测得 OpenSCAD 28.0%,而这三次连测是 34–36%。所以可比的条件不是"同一个评测脚本",而是**同一次会话内连续跑完**;隔了几小时再测,6–8pp 的漂移是可能的。今后任何模型间比较都必须在一个批次里跑完。
diff --git a/finetune/scripts/queue.txt b/finetune/scripts/queue.txt
index 4b2c169b..3d47f99d 100644
--- a/finetune/scripts/queue.txt
+++ b/finetune/scripts/queue.txt
@@ -12,3 +12,6 @@ noise_profile|bash scripts/run_noise_profile.sh
27b_mm_eval|bash scripts/run_27b_mm_eval.sh
bal9b|bash scripts/run_9b_balanced.sh
stopdpo_train|bash scripts/run_stop_dpo_train.sh
+best_dual|bash scripts/run_best_dual_decode.sh
+27b_dual|bash scripts/run_27b_dual_decode.sh
+stopdpo_r2|bash scripts/run_stop_dpo.sh
diff --git a/finetune/scripts/run_27b_dual_decode.sh b/finetune/scripts/run_27b_dual_decode.sh
new file mode 100755
index 00000000..55f3add9
--- /dev/null
+++ b/finetune/scripts/run_27b_dual_decode.sh
@@ -0,0 +1,17 @@
+#!/bin/bash
+# 27B v2 is the project's other headline model and its dialect numbers are greedy-only, so by section 13 they
+# understate it the same way the 9B's did (base 27B scores 88% on OpenSCAD, the fine-tune reads 34%).
+set -uo pipefail; cd /wekafs/ict/hx_624/llm-ft
+RUN=runs/lf_qwen38_27b_lora_v2
+if [ ! -d $RUN/merged ]; then
+ [ -f configs/lf/export_27b_v2.yaml ] || { echo "[27bdual] no export config"; exit 0; }
+ scripts/lf_export.sh configs/lf/export_27b_v2.yaml > logs/export_27b_v2_dual.log 2>&1 \
+ || { echo "[27bdual] EXPORT FAILED: $(grep -iE 'error|OSError' logs/export_27b_v2_dual.log | tail -1 | cut -c1-160)"; exit 0; }
+fi
+# a 51 GB checkpoint does not fit one shared card; two-way tensor parallel with the Mamba-cache cap it needs
+G=${GPUS:-$(scripts/free_gpus.sh 45000 | cut -d, -f1,2)}
+TP=$(echo "$G" | awk -F, '{print NF}')
+{ [ -z "$G" ] || [ "${TP:-0}" -lt 2 ]; } && { echo "[27bdual] skipped: need two cards with 45 GB free, got [$G]"; exit 0; }
+echo "[27bdual] $(date) on GPUs $G (tp=$TP)"
+GPUS=$G TP=$TP bash eval/eval_all_run2.sh $RUN/merged q27b_v2_t07 --temp 0.7 --seed 1 2>&1 | tail -10
+echo "[27bdual] DONE"
diff --git a/finetune/scripts/run_9b_full_dgx01.sh b/finetune/scripts/run_9b_full_dgx01.sh
new file mode 100755
index 00000000..10d78de6
--- /dev/null
+++ b/finetune/scripts/run_9b_full_dgx01.sh
@@ -0,0 +1,37 @@
+#!/bin/bash
+# The proper 9B run, on the seven empty cards of dgx01 (dgx03 has every card shared and has been the source of
+# most of the OOMs). Recipe = mmmix2's, which is the best SFT so far, with 34% more distinct code: every dialect,
+# one caption per sample (caption augmentation is already known not to add anything), plus the image sets.
+# Scored under BOTH decodings, because greedy alone is what hid the runaway problem for the whole project.
+set -uo pipefail; cd /wekafs/ict/hx_624/llm-ft
+RUN=runs/lf_qwen35_9b_full
+GPUS_USE=${GPUS:-0,1,2,3,4,5,7}
+echo "[full9b] $(date) train on dgx01 GPUs $GPUS_USE"
+GPUS=$GPUS_USE scripts/lf_train.sh configs/lf/lora_9b_full.yaml > logs/train_9b_full.log 2>&1
+[ -f $RUN/train_results.json ] || { echo "[full9b] TRAIN FAILED: $(grep -iE 'out of memory|Error|Traceback' logs/train_9b_full.log | grep -v errors | tail -1 | cut -c1-190)"; exit 1; }
+echo "[full9b] $(date) train done: $(tr -d '\n' < $RUN/train_results.json | cut -c1-150)"
+sed -e "s|model_name_or_path:.*|model_name_or_path: /wekafs/ict/hx_624/models/Qwen3.5-9B|" -e "s|adapter_name_or_path:.*|adapter_name_or_path: $(readlink -f $RUN)|" -e "s|export_dir:.*|export_dir: $(readlink -f $RUN)/merged|" configs/lf/export_lora_template.yaml > configs/lf/export_9b_full.yaml
+scripts/lf_export.sh configs/lf/export_9b_full.yaml > logs/export_9b_full.log 2>&1 || { echo "[full9b] EXPORT FAILED"; exit 1; }
+echo "[full9b] $(date) exported"
+G1=$(echo $GPUS_USE | cut -d, -f1)
+GPUS=$G1 TP=1 bash eval/eval_all_run2.sh $RUN/merged q9b_full_greedy 2>&1 | tail -10
+GPUS=$G1 TP=1 bash eval/eval_all_run2.sh $RUN/merged q9b_full_t07 --temp 0.7 --seed 1 2>&1 | tail -10
+source /wekafs/ict/hx_624/anaconda3/etc/profile.d/conda.sh; conda activate llmft
+python - <<'PY'
+import json, os
+suites = ["bench", "blender", "cadquery", "openscad", "glsl", "threejs"]
+sets = [("mmmix2 greedy", "q9b_mmmix2_all"), ("mmmix2 T=0.7", "q9b_mmmix2_t07"),
+ ("full9b greedy", "q9b_full_greedy"), ("full9b T=0.7", "q9b_full_t07")]
+for title, key in (("exec rate", "exec_rate"), ("mean output tokens", "mean_new_tokens")):
+ print(f" --- {title} ---")
+ print(f" {'model':16s}" + "".join(f"{s[:8]:>10s}" for s in suites))
+ for label, n in sets:
+ p = f"eval/out/{n}_report.json"
+ if not os.path.exists(p): print(f" {label:16s} (missing)"); continue
+ d = {x["suite"]: x for x in json.load(open(p))["suites"]}
+ if key == "exec_rate":
+ print(f" {label:16s}" + "".join(f"{d[s][key]*100:9.1f}%" if s in d else f"{'-':>10s}" for s in suites))
+ else:
+ print(f" {label:16s}" + "".join(f"{d[s].get(key,0):10,.0f}" if s in d else f"{'-':>10s}" for s in suites))
+PY
+echo "[full9b] ALL DONE"
diff --git a/finetune/scripts/run_best_dual_decode.sh b/finetune/scripts/run_best_dual_decode.sh
new file mode 100755
index 00000000..9c790916
--- /dev/null
+++ b/finetune/scripts/run_best_dual_decode.sh
@@ -0,0 +1,38 @@
+#!/bin/bash
+# Section 13 showed greedy understates every long-output dialect. The project's headline model was scored greedy
+# only, so its OpenSCAD/GLSL/three.js numbers are very likely wrong in the same direction. Re-score it under both
+# decodings so the report states what the best model can actually do, not what greedy lets it show.
+set -uo pipefail; cd /wekafs/ict/hx_624/llm-ft
+RUN=runs/lf_qwen35_9b_dpo_exec_v2
+# the DPO scripts delete their merged output to save disk, so the best model has to be rebuilt from its chain:
+# lora_v1/merged -> dpo_exec_v1/merged_r2 -> dpo_exec_v2/merged
+if [ ! -d $RUN/merged ]; then
+ if [ ! -d runs/lf_qwen35_9b_dpo_exec_v1/merged_r2 ]; then
+ echo "[bestdual] rebuilding dpo_exec_v1/merged_r2"
+ scripts/lf_export.sh configs/lf/export_dpo_exec_v1_r2.yaml > logs/export_best_r2.log 2>&1 \
+ || { echo "[bestdual] STAGE1 EXPORT FAILED: $(grep -iE 'error|OSError' logs/export_best_r2.log | tail -1 | cut -c1-160)"; exit 0; }
+ fi
+ echo "[bestdual] rebuilding dpo_exec_v2/merged"
+ scripts/lf_export.sh configs/lf/export_dpo_exec_v2.yaml > logs/export_best_dual.log 2>&1 \
+ || { echo "[bestdual] STAGE2 EXPORT FAILED: $(grep -iE 'error|OSError' logs/export_best_dual.log | tail -1 | cut -c1-160)"; exit 0; }
+fi
+G=""; for _ in $(seq 1 60); do G=${GPUS:-$(scripts/free_gpus.sh 40000 | cut -d, -f1)}; [ -n "$G" ] && break; sleep 30; done
+[ -z "$G" ] && { echo "[bestdual] no GPU with 40 GB free"; exit 0; }
+echo "[bestdual] $(date) on GPU $G"
+GPUS=$G TP=1 bash eval/eval_all_run2.sh $RUN/merged q9b_best_t07 --temp 0.7 --seed 1 2>&1 | tail -10
+source /wekafs/ict/hx_624/anaconda3/etc/profile.d/conda.sh; conda activate llmft
+python - <<'PY'
+import json, os
+suites = ["bench", "blender", "cadquery", "openscad", "glsl", "threejs"]
+for label, n in [("dpo_exec_v2 greedy", "qwen35_9b_dpo_exec_v2"), ("dpo_exec_v2 T=0.7", "q9b_best_t07")]:
+ p = f"eval/out/{n}_report.json"
+ if not os.path.exists(p):
+ s = f"eval/out/{n}/summary.json"
+ if os.path.exists(s):
+ d = json.load(open(s)); print(f" {label:20s} bench only: {d['exec_ok']}/{d['n_tasks']} = {d['exec_ok_rate']*100:.1f}%")
+ else: print(f" {label:20s} (no report)")
+ continue
+ d = {x["suite"]: x for x in json.load(open(p))["suites"]}
+ print(f" {label:20s}" + "".join(f"{d[s]['exec_rate']*100:9.1f}%" if s in d else f"{'-':>10s}" for s in suites))
+PY
+echo "[bestdual] DONE"