NOTE.md
July 22, 2025 ยท View on GitHub
Thanks for your interest in VisionThink.
On this page, we list the main modified files for your convenience in checking and integrating into your project.
๐ Dataset
verl/utils/dataset/multimodal_dataset.py
๐ Rollout
-
verl/trainer/ppo/ray_trainer.py
Main logic for training. -
verl/workers/rollout/vllm_rollout/vllm_rollout_spmd.py
Multi-turn rollout logic.
๐ง Reward Manager
-
verl/workers/reward_manager/naive_multithreads.py
GPT-based judge for General VQA. -
verl/workers/reward_manager/naive_multithreads_tool.py
GPT-based judge for General VQA and Efficient Reasoning VLM. -
verl/workers/reward_manager/verifier_reward_manager.py
Open-source LLM (e.g., Qwen) as judge.
๐ฏ Reward Function
-
verl/utils/reward_score/general_qa_gpt.py
GPT-based scoring for General VQA. -
verl/utils/reward_score/general_qa_tool.py
GPT-based scoring for General VQA and Efficient Reasoning VLM. -
verl/workers/reward_model/verifier/verifier.py
Scoring with open-source verifier models like Qwen.