Aletheia-Bench/GRPO-Think-7B-8k
Aletheia-Bench/GRPO-Think-7B-8k is a 7.6 billion parameter GRPO-Think code verifier model developed by Aletheia-Bench. Fine-tuned from DeepSeek-R1-Distill-Qwen-7B, it is trained with Reinforcement Learning with Verifiable Rewards (RLVR) using on-policy learning, negative samples, and intermediate thinking traces. This model specializes in judging and ranking candidate solutions for competitive programming problems, utilizing an 8k reasoning-token budget during training. It is designed for tasks requiring robust code verification and automated evaluation.
Loading preview...
Aletheia-Bench/GRPO-Think-7B-8k: A Code Verifier Model
GRPO-Think-7B-8k is a 7.6 billion parameter code verifier developed by Aletheia-Bench, fine-tuned from deepseek-ai/DeepSeek-R1-Distill-Qwen-7B. This model is part of the Aletheia project, which investigates the effectiveness of Reinforcement Learning with Verifiable Rewards (RLVR) for code verifiers. It is specifically trained using the GRPO (Generative Reinforcement Policy Optimization) approach, incorporating on-policy learning, negative samples, and intermediate thinking traces.
Key Capabilities and Features
- Code Verification: Designed to judge and rank candidate solutions for competitive programming problems, identifying the correct solution among several options.
- Thinking Traces: Utilizes intermediate thinking traces during training, allowing it to reason through code solutions with an 8k reasoning-token budget.
- RLVR Training: Benefits from a robust RLVR pipeline, leveraging both positive and negative samples for improved accuracy in ranking and selection.
- Foundation Model: Built upon the strong base of DeepSeek-R1-Distill-Qwen-7B, enhancing its code understanding and reasoning abilities.
Ideal Use Cases
- RLHF / RLAIF: Can serve as a plug-and-play reward function for optimizing code generation policies in Reinforcement Learning from Human/AI Feedback setups.
- Automated Evaluation: Suitable for LLM-as-a-judge applications across various code-related tasks, providing objective assessments of code quality and correctness.
- Research: Valuable for researchers studying the impact of thinking traces, on-policy learning, and negative samples in the development of effective code verifiers.