zhongweixie/qwen3vl-8b-claw-stage3-v2-lora

qwen3vl-8b-claw-stage3-v2-lora

LoRA adapter for Qwen3-VL-8B fine-tuned on claw-eval (stage3_v2). Best checkpoint: 0.506 total (C=0.778, T=0.619, M=0.224).

Model Details

• —Type: LoRA Adapter
• —Base Model: Qwen/Qwen3-VL-8B-Instruct
• —Training Method: LoRA (rank=32, alpha=64)
• —Training Data: sftmessagesstage3_v2.jsonl (included in this repo)
• —Evaluation: 300 tasks from claw-eval benchmark

Evaluation Results

Evaluated on claw-eval (300 tasks across C/T/M series):

Baseline scores: total=0.420, C=0.637, T=0.547, M=0.206

Usage

This is a LoRA adapter, not a full model. You need to load it with the base model:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-VL-8B-Instruct",
    torch_dtype="auto",
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "zhongweixie/qwen3vl-8b-claw-stage3-v2-lora")
tokenizer = AutoTokenizer.from_pretrained("zhongweixie/qwen3vl-8b-claw-stage3-v2-lora")

# Merge if you want a standalone model (optional)
# model = model.merge_and_unload()

# Your inference code here

Quick inference (using merge script from this repo):

# Download LoRA adapter
git clone https://huggingface.co/zhongweixie/qwen3vl-8b-claw-stage3-v2-lora

# Merge with base model
python train/merge_lora.py \
  --adapter qwen3vl-8b-claw-stage3-v2-lora \
  --output merged_model

# Serve
python train/serve_sft.py --model merged_model --port 8000

Training

See launch_stage3_v2.sh in this repo for the exact training command.

Key hyperparameters:

• —LoRA rank: 32, alpha: 64
• —Learning rate: 1e-6 (stage3_v2) / 1e-5 (v12)
• —Max sequence length: 8192
• —Epochs: 1

Files

• —*.safetensors: Model weights (merged LoRA + base)
• —sft_messages_stage3_v2.jsonl: Training data (SFT messages format)
• —launch_stage3_v2.sh: Training script
• —eval_results.json: Full evaluation breakdown

Citation

@misc{qwen3vl-claw-stage3_v2,
  author = {Zhongwei Xie},
  title = {Qwen3-VL-8B fine-tuned on claw-eval (stage3_v2)},
  year = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{zhongweixie/qwen3vl-8b-claw-stage3-v2-lora}},
}

License

Apache 2.0 (inherited from Qwen3-VL-8B-Instruct base model)