game24-rl
This repository contains two checkpoints from the game24-rl course project:
sft-final/: full fine-tuned Qwen/Qwen2.5-1.5B-Instruct checkpoint for 24-game reasoning.grpo-lora-final/: LoRA adapter trained from the SFT checkpoint with GRPO.
Training Route
Qwen2.5-1.5B-Instruct -> full SFT final -> GRPO LoRA adapter -> decoding/eval
Notes
The project focuses on standard 24-point game solving. The model is expected to produce reasoning and a final answer expression that can be checked by the repository verifier.
The GRPO artifact is a PEFT LoRA adapter, not a standalone full model. Load the SFT checkpoint first, then apply the adapter.
Reported Repo-Local Results
Under the repo-local train/validation/test split documented in the GitHub repository:
- Base model full-data direct eval:
16/1362 = 1.17%
- SFT final, validation, 1024 token budget:
110/136 = 80.88%
- GRPO LoRA, validation, 1024 token budget:
116/136 = 85.29%
- SFT final, validation, 4096 token budget:
123/136 = 90.44%
- SFT final, test, 4096 token budget:
128/137 = 93.43%
- GRPO LoRA, validation, 4096 token budget:
126/136 = 92.65%
- GRPO LoRA, test, 4096 token budget:
129/137 = 94.16%
See the GitHub handoff and experiment docs for split details, verifier details, decoding settings, and caveats.
