Official Codebase of 3DCodeVerse
Python
finetune: md_xl/md_bal builders, geometry-feedback DPO (geo_dpo_build.sh), exploration schedulers, ntok cache, GLB renderer + contact sheets; REPORT §9–10 (md_xl scale-up, curve, geometry DPO, re-balance)
gaoypengce40be067e28
diff --git a/finetune/configs/lf/dpo_geo_md_xl.yaml b/finetune/configs/lf/dpo_geo_md_xl.yamlnew file mode 100644index 00000000..9bd44edf--- /dev/null+++ b/finetune/configs/lf/dpo_geo_md_xl.yaml@@ -0,0 +1,42 @@+### DPO with execution-feedback preference pairs, on top of LoRA-v1 (merged)+model_name_or_path: /wekafs/ict/hx_624/llm-ft/runs/lf_qwen35_9b_lora_md_xl/merged_geo+trust_remote_code: true+flash_attn: fa2+stage: dpo+do_train: true+finetuning_type: lora+lora_rank: 16+lora_alpha: 32+lora_dropout: 0.05+lora_target: all+freeze_vision_tower: true+freeze_multi_modal_projector: true+pref_beta: 0.1+pref_loss: sigmoid+dataset_dir: /wekafs/ict/hx_624/llm-ft/data/lf+dataset: dpo_geo_pairs+template: qwen3_5+enable_thinking: false+cutoff_len: 8192+preprocessing_num_workers: 16+dataloader_num_workers: 2+output_dir: /wekafs/ict/hx_624/llm-ft/runs/lf_qwen35_9b_dpo_geo_md_xl+logging_steps: 5+save_steps: 1000+save_only_model: true+plot_loss: true+overwrite_output_dir: true+report_to: tensorboard+per_device_train_batch_size: 1+gradient_accumulation_steps: 8+learning_rate: 5.0e-6+num_train_epochs: 2.0+lr_scheduler_type: cosine+warmup_ratio: 0.1+bf16: true+gradient_checkpointing: true+ddp_timeout: 180000000+val_size: 0.02+per_device_eval_batch_size: 1+eval_strategy: steps+eval_steps: 100diff --git a/finetune/configs/lf/dpo_md_xl.yaml b/finetune/configs/lf/dpo_md_xl.yamlnew file mode 100644index 00000000..7b3397ad--- /dev/null+++ b/finetune/configs/lf/dpo_md_xl.yaml@@ -0,0 +1,42 @@+### DPO with execution-feedback preference pairs, on top of LoRA-v1 (merged)+model_name_or_path: /wekafs/ict/hx_624/llm-ft/runs/lf_qwen35_9b_lora_md_xl/merged+trust_remote_code: true+flash_attn: fa2+stage: dpo+do_train: true+finetuning_type: lora+lora_rank: 16+lora_alpha: 32+lora_dropout: 0.05+lora_target: all+freeze_vision_tower: true+freeze_multi_modal_projector: true+pref_beta: 0.1+pref_loss: sigmoid+dataset_dir: /wekafs/ict/hx_624/llm-ft/data/lf+dataset: dpo_md_xl_pairs+template: qwen3_5+enable_thinking: false+cutoff_len: 8192+preprocessing_num_workers: 16+dataloader_num_workers: 2+output_dir: /wekafs/ict/hx_624/llm-ft/runs/lf_qwen35_9b_dpo_md_xl+logging_steps: 5+save_steps: 1000+save_only_model: true+plot_loss: true+overwrite_output_dir: /wekafs/ict/hx_624/llm-ft/runs/lf_qwen35_9b_dpo_md_xl+report_to: tensorboard+per_device_train_batch_size: 1+gradient_accumulation_steps: 8+learning_rate: 5.0e-6+num_train_epochs: 2.0+lr_scheduler_type: cosine+warmup_ratio: 0.1+bf16: true+gradient_checkpointing: true+ddp_timeout: 180000000+val_size: 0.02+per_device_eval_batch_size: 1+eval_strategy: steps+eval_steps: 100diff --git a/finetune/configs/lf/export_dpo_geo_md_xl.yaml b/finetune/configs/lf/export_dpo_geo_md_xl.yamlnew file mode 100644index 00000000..ee105071--- /dev/null+++ b/finetune/configs/lf/export_dpo_geo_md_xl.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_lora_md_xl/merged_geo+adapter_name_or_path: /wekafs/ict/hx_624/llm-ft/runs/lf_qwen35_9b_dpo_geo_md_xl+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_dpo_geo_md_xl/merged+export_size: 5+export_device: cpu+export_legacy_format: falsediff --git a/finetune/configs/lf/export_dpo_md_xl.yaml b/finetune/configs/lf/export_dpo_md_xl.yamlnew file mode 100644index 00000000..bda73106--- /dev/null+++ b/finetune/configs/lf/export_dpo_md_xl.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_lora_md_xl/merged+adapter_name_or_path: /wekafs/ict/hx_624/llm-ft/runs/lf_qwen35_9b_dpo_md_xl+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_dpo_md_xl/merged+export_size: 5+export_device: cpu+export_legacy_format: falsediff --git a/finetune/configs/lf/export_md_bal.yaml b/finetune/configs/lf/export_md_bal.yamlnew file mode 100644index 00000000..1a9a7f8b--- /dev/null+++ b/finetune/configs/lf/export_md_bal.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_lora_md_bal+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_lora_md_bal/merged+export_size: 5+export_device: cpu+export_legacy_format: falsediff --git a/finetune/configs/lf/export_md_xl_ck1350.yaml b/finetune/configs/lf/export_md_xl_ck1350.yamlnew file mode 100644index 00000000..147064c7--- /dev/null+++ b/finetune/configs/lf/export_md_xl_ck1350.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_lora_md_xl/checkpoint-1350+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_lora_md_xl/checkpoint-1350/merged+export_size: 5+export_device: cpu+export_legacy_format: falsediff --git a/finetune/configs/lf/export_md_xl_ck1800.yaml b/finetune/configs/lf/export_md_xl_ck1800.yamlnew file mode 100644index 00000000..49c8bbf0--- /dev/null+++ b/finetune/configs/lf/export_md_xl_ck1800.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_lora_md_xl/checkpoint-1800+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_lora_md_xl/checkpoint-1800/merged+export_size: 5+export_device: cpu+export_legacy_format: falsediff --git a/finetune/configs/lf/export_md_xl_ck2250.yaml b/finetune/configs/lf/export_md_xl_ck2250.yamlnew file mode 100644index 00000000..238e03cf--- /dev/null+++ b/finetune/configs/lf/export_md_xl_ck2250.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_lora_md_xl/checkpoint-2250+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_lora_md_xl/checkpoint-2250/merged+export_size: 5+export_device: cpu+export_legacy_format: falsediff --git a/finetune/configs/lf/export_md_xl_ck2700.yaml b/finetune/configs/lf/export_md_xl_ck2700.yamlnew file mode 100644index 00000000..8d739754--- /dev/null+++ b/finetune/configs/lf/export_md_xl_ck2700.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_lora_md_xl/checkpoint-2700+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_lora_md_xl/checkpoint-2700/merged+export_size: 5+export_device: cpu+export_legacy_format: falsediff --git a/finetune/configs/lf/export_md_xl_ck3150.yaml b/finetune/configs/lf/export_md_xl_ck3150.yamlnew file mode 100644index 00000000..838a7dc1--- /dev/null+++ b/finetune/configs/lf/export_md_xl_ck3150.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_lora_md_xl/checkpoint-3150+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_lora_md_xl/checkpoint-3150/merged+export_size: 5+export_device: cpu+export_legacy_format: falsediff --git a/finetune/configs/lf/export_md_xl_ck3600.yaml b/finetune/configs/lf/export_md_xl_ck3600.yamlnew file mode 100644index 00000000..8009ea3b--- /dev/null+++ b/finetune/configs/lf/export_md_xl_ck3600.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_lora_md_xl/checkpoint-3600+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_lora_md_xl/checkpoint-3600/merged+export_size: 5+export_device: cpu+export_legacy_format: falsediff --git a/finetune/configs/lf/export_md_xl_ck900.yaml b/finetune/configs/lf/export_md_xl_ck900.yamlnew file mode 100644index 00000000..21c131e3--- /dev/null+++ b/finetune/configs/lf/export_md_xl_ck900.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_lora_md_xl/checkpoint-900+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_lora_md_xl/checkpoint-900/merged+export_size: 5+export_device: cpu+export_legacy_format: falsediff --git a/finetune/data_specs/dataset_info.json b/finetune/data_specs/dataset_info.jsonindex d4cb7d10..2611010d 100644--- a/finetune/data_specs/dataset_info.json+++ b/finetune/data_specs/dataset_info.json@@ -346,5 +346,27 @@"messages": "conversations","system": "system"}+ },+ "dpo_md_xl_pairs": {+ "file_name": "dpo_md_xl_pairs.json",+ "ranking": true,+ "formatting": "sharegpt",+ "columns": {+ "messages": "conversations",+ "system": "system",+ "chosen": "chosen",+ "rejected": "rejected"+ }+ },+ "dpo_geo_pairs": {+ "file_name": "dpo_geo_pairs.json",+ "ranking": true,+ "formatting": "sharegpt",+ "columns": {+ "messages": "conversations",+ "system": "system",+ "chosen": "chosen",+ "rejected": "rejected"+ }}}\ No newline at end of filediff --git a/finetune/docs/assets/md_xl_vs_geo_dpo_contact_sheet.png b/finetune/docs/assets/md_xl_vs_geo_dpo_contact_sheet.pngnew file mode 100644index 00000000..8cfe8c68Binary files /dev/null and b/finetune/docs/assets/md_xl_vs_geo_dpo_contact_sheet.png differdiff --git a/finetune/scripts/contact_sheet_compare.py b/finetune/scripts/contact_sheet_compare.pynew file mode 100644index 00000000..750836b2--- /dev/null+++ b/finetune/scripts/contact_sheet_compare.py@@ -0,0 +1,32 @@+"""3-column contact sheet on the SAME tasks: model A vs model B vs GT (tasks = the selection used by contact_sheet_3dcodebench.py for model A)."""+import json, os, sys, subprocess, random+from concurrent.futures import ThreadPoolExecutor+from PIL import Image, ImageDraw, ImageFont+BL = "/wekafs/ict/hx_624/tools/blender-5.0.1-linux-x64/blender"; R = os.path.abspath("eval/render_glb.py"); GT = "/wekafs/ict/hx_624/data/3dcodebench/data"+dirA, nameA, dirB, nameB, out_png = sys.argv[1:6]; k = 4+mA = {r["task"]: r for r in map(json.loads, open(f"{dirA}/metrics.jsonl"))}; mB = {r["task"]: r for r in map(json.loads, open(f"{dirB}/metrics.jsonl"))}+ok = sorted([r for r in mA.values() if r["exec"] == "OK" and r.get("[email protected]") is not None], key=lambda r: -r["[email protected]"]); random.seed(3); n = len(ok)+sel = [("best", r) for r in ok[:k]] + [("median", r) for r in ok[n//2 - k//2: n//2 - k//2 + k]] + [("worst", r) for r in ok[-k:]] + [("random", r) for r in random.sample(ok[k:-k], k)]+def render(glb, png):+ if os.path.exists(png) or not os.path.exists(glb): return+ subprocess.run([BL, "-b", "--factory-startup", "-noaudio", "--python", R, "--", "--glb", glb, "--out", png], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, timeout=300, env={**os.environ, "HOME": "/tmp"})+wA = os.path.abspath(f"{dirA}/_renders"); wB = os.path.abspath(f"{dirB}/_renders"); os.makedirs(wA, exist_ok=True); os.makedirs(wB, exist_ok=True)+jobs = []+for tag, r in sel:+ t = r["task"]; jobs += [(os.path.abspath(f"{dirA}/{t}/out.glb"), f"{wA}/{t}_gen.png"), (os.path.abspath(f"{dirB}/{t}/out.glb"), f"{wB}/{t}_gen.png"), (f"{GT}/{t}/glb/{t}.glb", f"{wA}/{t}_gt.png")]+with ThreadPoolExecutor(8) as ex: list(ex.map(lambda j: render(*j), jobs))+S = 180; cols = k; rows = 4+sheet = Image.new("RGB", (cols * (3 * S + 16) + 40, rows * (S + 52) + 10), "white"); d = ImageDraw.Draw(sheet)+try: font = ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf", 12); fb = ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf", 12)+except Exception: font = fb = ImageFont.load_default()+for i, (tag, r) in enumerate(sel):+ row, col = i // cols, i % cols; x0 = 40 + col * (3 * S + 16); y0 = 10 + row * (S + 52); t = r["task"]+ for j, p in enumerate((f"{wA}/{t}_gen.png", f"{wB}/{t}_gen.png", f"{wA}/{t}_gt.png")):+ try: im = Image.open(p).convert("RGB").resize((S, S))+ except Exception: im = Image.new("RGB", (S, S), (230, 230, 230))+ sheet.paste(im, (x0 + j * S, y0 + 20))+ fa = r["[email protected]"]; fbv = mB.get(t, {}).get("[email protected]"); sb = f"{fbv:.2f}" if fbv is not None else mB.get(t, {}).get("exec", "?")+ d.text((x0, y0 + 2), f"{t.replace('_seed0','')} [email protected] {nameA}={fa:.2f} | {nameB}={sb}", fill="black", font=fb)+ d.text((x0, y0 + S + 24), nameA, fill=(60, 60, 60), font=font); d.text((x0 + S, y0 + S + 24), nameB, fill=(60, 60, 60), font=font); d.text((x0 + 2 * S, y0 + S + 24), "ground truth", fill=(60, 60, 60), font=font)+ if col == 0: d.text((4, y0 + S // 2), tag, fill=(120, 0, 0), font=fb)+sheet.save(out_png); print("wrote", out_png, sheet.size)diff --git a/finetune/scripts/explore_sched2.sh b/finetune/scripts/explore_sched2.shnew file mode 100755index 00000000..055bf300--- /dev/null+++ b/finetune/scripts/explore_sched2.sh@@ -0,0 +1,36 @@+#!/bin/bash+# revised 8h scheduler (GPUs 0,2,3 idle now, GPU 1 busy with the md_xl sweep):+# (1) now: md_bal LoRA on GPUs 0,2,3 (train+merge+evals) (2) after sweep: geometry-feedback DPO on GPU 1 (train+export+evals)+# (3) after both: full FT md_bal on GPUs 0,1,2,3 (ZeRO-3) + evals+set -uo pipefail; cd /wekafs/ict/hx_624/llm-ft+echo "[sched] $(date) md_bal LoRA on GPUs 0,2,3"+scripts/run_md_exp.sh 0,2,3 md_bal > logs/exp_md_bal.log 2>&1 &+( until grep -q "\[bigsweep\] DONE" logs/md_xl_ckpt_sweep.log 2>/dev/null; do sleep 60; done+ echo "[sched] $(date) geometry-feedback DPO on GPU 1"+ GPUS=1 scripts/lf_train.sh configs/lf/dpo_geo_md_xl.yaml gradient_accumulation_steps=8 cutoff_len=4096 > logs/train_dpo_geo_md_xl.log 2>&1+ RUN=runs/lf_qwen35_9b_dpo_geo_md_xl+ if [ ! -f $RUN/train_results.json ]; then echo "[sched] GEO DPO TRAIN FAILED"; grep -E "Error" logs/train_dpo_geo_md_xl.log | grep -v errors | tail -2 | cut -c1-200; echo "[sched] GEO_DONE"; exit 0; fi+ echo "[sched] $(date) geo dpo train done: $(tr -d '\n' < $RUN/train_results.json | cut -c1-200)"+ BASE=runs/lf_qwen35_9b_lora_md_xl/merged_geo+ sed -e "s|model_name_or_path:.*|model_name_or_path: $(readlink -f $BASE)|" -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_dpo_geo_md_xl.yaml+ scripts/lf_export.sh configs/lf/export_dpo_geo_md_xl.yaml > logs/export_dpo_geo_md_xl.log 2>&1+ GPUS=1 eval/run_eval_vllm.sh $RUN/merged eval/out/qwen35_9b_dpo_geo_md_xl --no_think --max_new_tokens 6144 > logs/eval_dpo_geo_md_xl.log 2>&1+ echo "[sched] geo dpo 3DCodeBench: $(python3 -c "import json; s=json.load(open('eval/out/qwen35_9b_dpo_geo_md_xl/summary.json')); print(s['exec_ok'], round(s['[email protected]_mean_all(fail=0)'],3), 'scored', round(s['[email protected]_mean_scored'],3), '[email protected] scored', round(s['[email protected]_mean_scored'],3))")"+ GPUS=1 eval/run_dialect_eval.sh $RUN/merged dpo_geo_md_xl blender cadquery > logs/mdeval_dpo_geo_md_xl.log 2>&1+ for D in blender cadquery; do echo "[sched] geo dpo $D: $(cat eval/out/md_dpo_geo_md_xl_$D/summary.json | tr -d '\n ' | cut -c1-200)"; done+ GPUS=1 eval/best_of_n.sh $RUN/merged dpo_geo_md_xl_T07 4 0.7 2>&1 | grep -E "pass@" | tail -1 | sed 's/^/[sched] geo dpo /'+ echo "[sched] GEO_DONE" ) &+until grep -qE "\[md md_bal\] (DONE|TRAIN FAILED|EXPORT FAILED)" logs/exp_md_bal.log 2>/dev/null; do sleep 60; done+until grep -q "GEO_DONE" logs/explore_sched2.log 2>/dev/null; do sleep 60; done+echo "[sched] $(date) full FT md_bal on GPUs 0,1,2,3 (ZeRO-3, accum 2)"+GPUS=0,1,2,3 scripts/lf_train.sh configs/lf/full_md_bal.yaml gradient_accumulation_steps=2 > logs/train_full_md_bal.log 2>&1+RUN=runs/lf_qwen35_9b_full_md_bal+if [ -f $RUN/train_results.json ]; then+ echo "[sched] $(date) full md_bal train done: $(tr -d '\n' < $RUN/train_results.json | cut -c1-200)"+ GPUS=0 eval/run_eval_vllm.sh $RUN eval/out/qwen35_9b_full_md_bal --no_think --max_new_tokens 6144 > logs/eval_full_md_bal.log 2>&1 &+ GPUS=1 eval/run_dialect_eval.sh $RUN full_md_bal > logs/mdeval_full_md_bal.log 2>&1 &+ wait+ echo "[sched] full md_bal 3DCodeBench: $(python3 -c "import json; s=json.load(open('eval/out/qwen35_9b_full_md_bal/summary.json')); print(s['exec_ok'], round(s['[email protected]_mean_all(fail=0)'],3))")"+ for D in cadquery openscad glsl blender; do echo "[sched] full md_bal $D: $(cat eval/out/md_full_md_bal_$D/summary.json | tr -d '\n ' | cut -c1-160)"; done+else echo "[sched] FULL TRAIN FAILED"; grep -E "Error" logs/train_full_md_bal.log | grep -v errors | tail -2 | cut -c1-200; fi+echo "[sched] FULL_DONE"; echo "[sched] ALL DONE"