Aletheia-Bench/GRPO-Think-7B-4k
Aletheia-Bench/GRPO-Think-7B-4k is a 7.6 billion parameter GRPO-Think verifier model developed by Aletheia-Bench, fine-tuned from DeepSeek-R1-Distill-Qwen-7B. It is specifically designed for judging and ranking candidate code solutions for competitive programming problems, utilizing a 4k reasoning-token budget during training. This model excels as a plug-and-play reward function for code generation policy optimization and automated evaluation.
Loading preview...
Aletheia-Bench/GRPO-Think-7B-4k Overview
This model is a GRPO-Think verifier, a specialized thinking code verifier developed by Aletheia-Bench. It is fine-tuned from the 7B-scale deepseek-ai/DeepSeek-R1-Distill-Qwen-7B and trained using GRPO (on-policy, with negative samples and intermediate thinking traces) with a 4k reasoning-token budget for maximum completion length during training.
Key Capabilities
- Code Verification: Judges and ranks candidate solutions for competitive programming problems, given a problem statement and multiple code options.
- RLVR Integration: Designed to be a plug-and-play reward function for Reinforcement Learning with Verifiable Rewards (RLVR) in code generation policy optimization.
- Automated Evaluation: Functions as an LLM-as-a-judge for various code-related tasks.
- Research Tool: Useful for studying the impact of thinking traces, on-policy learning, and negative samples in training robust code verifiers.
Training Details
The model was trained on the Aletheia-Bench/Aletheia-Train dataset, which comprises 50,000 training instances of competitive programming problems paired with 2-5 candidate solutions. The Aletheia dataset collection also includes variants for DPO training and adversarial testing, enabling comprehensive analysis of verifier performance.