cfcamo/cfcamo-sft-4b
The cfcamo/cfcamo-sft-4b is a 4 billion parameter vision-language model, based on Qwen3-VL-4B-Instruct, developed by Suhang Li, Osamu Yoshie, and Yuya Ieiri. It serves as a cold-start checkpoint fine-tuned on 1000 paired SFT rows to teach a detect-or-abstain output schema, specifically for counterfactual camouflaged object detection. This model is designed as the initial stage for further Reinforcement Learning (RL) training, providing a base that already abstains on target-absent counterfactuals.
Loading preview...
CFCamo-SFT-4B: A Foundation for Counterfactual Object Detection
cfcamo/cfcamo-sft-4b is a 4 billion parameter vision-language model, initialized from Qwen3-VL-4B-Instruct. Developed by Suhang Li, Osamu Yoshie, and Yuya Ieiri, this model is a crucial cold-start checkpoint in the CFCamo framework, designed to teach a specific detect-or-abstain output schema.
Key Capabilities & Training
- Detect-or-Abstain Schema: Fine-tuned on 1000 paired Supervised Fine-Tuning (SFT) rows (500 detect + 500 abstain) to enable the model to abstain when a target object is absent in counterfactual scenarios.
- Base for RL: This SFT checkpoint is the initial stage from which subsequent Reinforcement Learning (RL) stages (LoRA or full fine-tuning) are trained, leading to the main performance numbers presented in the associated paper.
- Training Details: Trained for 1 epoch with a learning rate of 2e-5, using a batch size of 2 and gradient accumulation of 8.
Further Development & Resources
- Users can continue training this model with RL using the provided code and training recipe.
- The project includes a dedicated benchmark (CF-COD) and training data.
- The paper detailing the CFCamo framework is available on arXiv.
- Pre-trained RL checkpoints, built on top of this SFT base, are also available as cfcamo/cfcamo-rl-lora and cfcamo/cfcamo-rl-full.