cfcamo/cfcamo-rl-lora

CFCamo-4B (RL-LoRA adapter)

LoRA adapter for CFCamo: trained on top of cfcamo/cfcamo-sft-4b with Counterfactual Sequence Policy Optimization (CSPO) and a Counterfactual Paired Reward (CPR). This is the paper LoRA checkpoint at step 252 (ε=0.5) over the 4040-pair RL set — the single-large-GPU alternative to cfcamo/cfcamo-rl-full.

• 📄 Paper: https://arxiv.org/abs/2606.11231
• Code & training recipe: https://github.com/suhang2000/CFCamo
• Benchmark (CF-COD) + training data: https://huggingface.co/datasets/cfcamo/CF-COD

Use it (PEFT load)

from transformers import AutoModelForImageTextToText, AutoProcessor
from peft import PeftModel

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

Or merge into a single standalone HF model:

python scripts/eval/merge_lora.py \
  --base checkpoints/cfcamo-sft-4b \
  --lora checkpoints/cfcamo-rl-lora \
  --out checkpoints/cfcamo-rl-lora-merged

Reproduce paper LoRA numbers

git clone https://github.com/suhang2000/CFCamo && cd CFCamo
pip install -e ".[eval]"
huggingface-cli download cfcamo/cfcamo-sft-4b --local-dir checkpoints/cfcamo-sft-4b
huggingface-cli download cfcamo/cfcamo-rl-lora --local-dir checkpoints/cfcamo-rl-lora
huggingface-cli download --repo-type dataset cfcamo/CF-COD --local-dir data/cfcod
# (place upstream COD into data/cfcod/<source>/{Imgs,GT}/ — see dataset card)

python scripts/eval/merge_lora.py \
  --base checkpoints/cfcamo-sft-4b \
  --lora checkpoints/cfcamo-rl-lora \
  --out checkpoints/cfcamo-rl-lora-merged

python scripts/eval/eval_cfcod.py \
  --cf-manifest data/cfcod/test/cf_manifest_test.jsonl \
  --data-root data/cfcod \
  --models "CFCamo-LoRA=checkpoints/cfcamo-rl-lora-merged" \
  --out-dir results/cfcod_eval

Training summary

• Base: cfcamo/cfcamo-sft-4b (SFT cold-start on Qwen3-VL-4B-Instruct)
• RL: CSPO/CPR (eta=1.0), LoRA r=64, single large GPU
• Checkpoint at step 252 = ε=0.5 over the 4040-pair RL set

Citation

@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}
}