Use it

python
Copy code
from transformers import AutoModelForImageTextToText, AutoProcessor
from PIL import Image

ckpt = "cfcamo/cfcamo-sft-4b"
processor = AutoProcessor.from_pretrained(ckpt)
model = AutoModelForImageTextToText.from_pretrained(
    ckpt, torch_dtype="auto", device_map="auto",
).eval()
# (use the same detect-or-abstain prompt as cfcamo-rl-full — see that model card)

Continue training: RL on top

bash
Copy code
git clone https://github.com/suhang2000/CFCamo && cd CFCamo
pip install -e .
huggingface-cli download cfcamo/cfcamo-sft-4b --local-dir checkpoints/cfcamo-sft-4b
huggingface-cli download --repo-type dataset cfcamo/CF-COD --local-dir data/cfcod
# LoRA (single GPU) — paper-main LoRA checkpoint at step 252 (epsilon=0.5)
python -m verl.trainer.main config=configs/rl_lora.yaml
# Full fine-tuning (multi-GPU) — checkpoint at step 126 (epsilon=0.5)
python -m verl.trainer.main config=configs/rl_full.yaml

The two RL checkpoints are also released at cfcamo/cfcamo-rl-lora (LoRA adapter on top of this SFT base) and cfcamo/cfcamo-rl-full.

Training summary

• Base: Qwen/Qwen3-VL-4B-Instruct (Apache-2.0)
• Data: 1000 SFT rows (paired r=1:1) with paired CoT reasoning
• 1 epoch, lr 2e-5, batch=2 × grad_accum=8

Citation

bibtex
Copy code
@article{li2026cfcamo,
  title   = {{CFCamo}: A Counterfactual Detect-or-Abstain Framework for Camouflaged Object Detection},
  author  = {Li, Suhang and Yoshie, Osamu and Ieiri, Yuya},
  journal = {arXiv preprint arXiv:2606.11231},
  year    = {2026}
}