README.md

---
license: apache-2.0
base_model: OpenGVLab/InternVL2-8B
tags:
  - internvl2
  - cad
  - cadquery
  - code-generation
  - image-to-text
  - fine-tuned
datasets:
  - CADCODER/GenCAD-Code
language:
  - ja
pipeline_tag: image-text-to-text
---

# InternVL2-8B GenCAD-Code Fine-tuned (v2)

An InternVL2-8B model fine-tuned to generate CadQuery Python code from 3D CAD model images.

## Model Details

- **Base Model**: [OpenGVLab/InternVL2-8B](https://huggingface.co/OpenGVLab/InternVL2-8B)
- **Fine-tuning Data**: [CADCODER/GenCAD-Code](https://huggingface.co/datasets/CADCODER/GenCAD-Code) 10,000 samples (extracted from 147K)
- **Training**: Full fine-tuning (no LoRA), 3 epochs, best at epoch 2
- **Best Validation Loss**: 0.1487
- **Hardware**: 4x NVIDIA RTX 6000 Ada (48GB each)
- **Training Framework**: aiDaptive (Phison)

## Performance

50-sample evaluation (stratified sampling from eval.json, unseen during training):

| Metric | Base Model | Fine-tuned (v2) | Improvement |
|--------|:----------:|:---------------:|:-----------:|
| Average Loss | 1.166 | **0.192** | -83.5% |
| Python Syntax Valid | 34/50 | **47/50** | +38% |
| CadQuery Execution | 1/50 | **46/50** | +4500% |
| Solid Generation | 0/50 | **46/50** | - |
| STL Export | 0/50 | **46/50** | - |

## Training Configuration (v2)

| Parameter | Value |
|-----------|-------|
| Learning Rate | 5e-6 (Cosine Annealing) |
| Effective Batch Size | 32 (1/GPU × 4GPU × grad_accum=8) |
| Epochs | 3 |
| Max Sequence Length | 4096 |
| Weight Decay | 0.05 |
| Precision | bf16 |
| Early Stopping | patience=2, min_delta=0.005 |

## Usage

```python
import torch
from transformers import AutoTokenizer, AutoModel
from PIL import Image

model_path = "Nextorage/InternVL2-8B-GenCAD-Code-v2"
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModel.from_pretrained(
    model_path,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
).eval().cuda()

# System prompt
system_prompt = (
    "あなたはCADコード生成アシスタントです。"
    "3D CADモデルの画像が与えられた場合、そのモデルを再現する "
    "CadQuery Pythonコードを生成してください。"
    "説明は不要です。コードのみを出力してください。"
)
```

## Limitations

- Max token length: 4096 (very complex models may be truncated)
- Trained on GenCAD-Code dataset only (specific CadQuery patterns)
- Japanese system prompt used during training

## Citation

If you use this model, please cite the GenCAD-Code dataset:

```
@misc{gencad-code,
  title={GenCAD-Code Dataset},
  author={CADCODER},
  url={https://huggingface.co/datasets/CADCODER/GenCAD-Code}
}
```