Official Codebase of 3DCodeVerse
Python
mm: the three-way ablation says a reference image adds nothing over the caption
gaoypengf7858fa7fe17
diff --git a/finetune/configs/lf/export_mm_img.yaml b/finetune/configs/lf/export_mm_img.yamlnew file mode 100644index 00000000..99fe5dff--- /dev/null+++ b/finetune/configs/lf/export_mm_img.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_mm_img+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_mm_img/merged+export_size: 5+export_device: cpu+export_legacy_format: falsediff --git a/finetune/configs/lf/export_mm_text.yaml b/finetune/configs/lf/export_mm_text.yamlnew file mode 100644index 00000000..cd14ea17--- /dev/null+++ b/finetune/configs/lf/export_mm_text.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_mm_text+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_mm_text/merged+export_size: 5+export_device: cpu+export_legacy_format: falsediff --git a/finetune/configs/lf/lora_27b_mm_img_text.yaml b/finetune/configs/lf/lora_27b_mm_img_text.yamlindex 034da312..4e51f1db 100644--- a/finetune/configs/lf/lora_27b_mm_img_text.yaml+++ b/finetune/configs/lf/lora_27b_mm_img_text.yaml@@ -9,18 +9,19 @@ lora_alpha: 128lora_dropout: 0.05lora_target: allfreeze_vision_tower: true+image_max_pixels: 200704freeze_multi_modal_projector: truedataset_dir: /wekafs/ict/hx_624/llm-ft/data/mm1dataset: 3dcv_mm_img_text_traineval_dataset: 3dcv_mm_img_text_valtemplate: qwen3_5enable_thinking: false-cutoff_len: 6144+cutoff_len: 4096preprocessing_num_workers: 16dataloader_num_workers: 4neat_packing: falsepacking: false-tokenized_path: /wekafs/ict/hx_624/llm-ft/data/lf/tok_mm1_27b_img_text+tokenized_path: /wekafs/ict/hx_624/llm-ft/data/lf/tok_mm1_27b_img_text_len4096_px448output_dir: /wekafs/ict/hx_624/llm-ft/runs/lf_qwen38_27b_mm_img_textlogging_steps: 5save_steps: 300diff --git a/finetune/scripts/run_9b_mmmix2.sh b/finetune/scripts/run_9b_mmmix2.shindex ad2c5e69..fcc82eff 100755--- a/finetune/scripts/run_9b_mmmix2.sh+++ b/finetune/scripts/run_9b_mmmix2.sh@@ -2,7 +2,7 @@# Image-conditioned pilot: render -> code. The qwen3_5 template already carries the qwen3_vl multimodal plugin,# so this is the first run in the project that feeds the model what a caption cannot express — the target's shape.set -uo pipefail; cd /wekafs/ict/hx_624/llm-ft-G=$(scripts/free_gpus.sh); N=$(echo "$G" | awk -F, '{print NF}')+G=${GPUS:-$(scripts/free_gpus.sh)}; N=$(echo "$G" | awk -F, '{print NF}')echo "[9bmmmix2] $(date) smoke on GPUs $G"GPUS=$G scripts/lf_train.sh configs/lf/lora_9b_mmmix2.yaml max_steps=6 save_steps=100000 eval_steps=100000 \output_dir=/wekafs/ict/hx_624/llm-ft/runs/_smoke_mmmix24 > logs/smoke_mmmix2.log 2>&1diff --git a/finetune/scripts/run_9b_mmvis.sh b/finetune/scripts/run_9b_mmvis.shindex 642ee235..0a63522a 100755--- a/finetune/scripts/run_9b_mmvis.sh+++ b/finetune/scripts/run_9b_mmvis.sh@@ -2,7 +2,7 @@# Image-conditioned pilot: render -> code. The qwen3_5 template already carries the qwen3_vl multimodal plugin,# so this is the first run in the project that feeds the model what a caption cannot express — the target's shape.set -uo pipefail; cd /wekafs/ict/hx_624/llm-ft-G=$(scripts/free_gpus.sh); N=$(echo "$G" | awk -F, '{print NF}')+G=${GPUS:-$(scripts/free_gpus.sh)}; N=$(echo "$G" | awk -F, '{print NF}')echo "[9bmmvis] $(date) smoke on GPUs $G"GPUS=$G scripts/lf_train.sh configs/lf/lora_9b_mmvis.yaml max_steps=6 save_steps=100000 eval_steps=100000 \output_dir=/wekafs/ict/hx_624/llm-ft/runs/_smoke_mmvis4 > logs/smoke_mmvis.log 2>&1