Fine‑tune Qwen 2.5‑VL (or any Vision‑Language model with the same API) on image grounding tasks using GRPO (Generic Reward Prediction Optimization) in just a few lines of code.

• Plug‑and‑play trainer – drop in your own JSON dataset of prompts + bounding‑boxes and start training.
• Image‑aware data collator – automatically loads, preprocesses and batches images.
• Reward‑based optimisation – leverages the trl library’s GRPO algorithm for RL‑style fine‑tuning.
• Minimal codebase – only three Python files, easy to read and customise.

• Accepts an image_processor and an images_root folder.
• Overrides data_collator to
  1. Load images with Pillow.
  2. Batch‑encode them via the Hugging Face AutoProcessor.
  3. Return a dict containing
     • pixel_values – tensor (C × H × W)
     • prompt – instruction string
     • solution – ground‑truth bbox or coordinates
     • scales – original image size

Tiny subclass that forwards all arguments to the real Qwen 2.5‑VL model while gracefully ignoring the extra logits_to_keep parameter expected by GRPO.

Currently only accuracy_reward_coord, which returns 1 if the (x, y) coordinate predicted by the model falls inside the ground‑truth bounding‑box and 0 otherwise.
Feel free to add IoU‑ or distance‑based rewards here.

Provides a concrete example wiring everything together.
Customise the constants at the top, or replace them with argparse flags for production use.

Released under the MIT License – free to use, modify and distribute.

• TRL library for GRPO.
• Qwen‑VL team for the open‑source model.