Fix CUDA device-side assert by adjusting max_prompt_length and disabling use_cache

Fix IndentationError and duplicate reward logic

Fix dtype mismatch in training and update blog

fix: resolve tensor size mismatch in reward_function by indexing correctly

fix: resolve Gradio JSON error by passing dicts directly instead of serialized strings

fix: resolve 422 errors by sanitizing JSON and fix python syntax error

fix: explicitly add unsloth_zoo to requirements to resolve ImportError

fix: unpin torch and use simpler unsloth install to resolve HF build cache miss

fix: downgrade CUDA to 12.1 and pin torch versions for HF Space compatibility

fix: improve GRPO learning signal and handle 422 environment errors

feat: randomize number of facts (2-4) per episode for better training diversity

fix: address reward plateau by adding format rewards and improving GRPO logic

fix: resolve GPU detection issue on HF Spaces

Update app.py

Update requirements.txt

Update requirements.txt

Update requirements.txt

Update requirements.txt

Create requirements.txt

Rename ap.py to app.py

Rename train_suspect_x.py to ap.py

Upload train_suspect_x.py with huggingface_hub

initial commit