Aletheia-Bench/GRPO-Instruct-14B
Aletheia-Bench/GRPO-Instruct-14B is a 14.8 billion parameter instruction-tuned code verifier developed by Aletheia-Bench, fine-tuned from Qwen/Qwen2.5-14B-Instruct. This model is trained using Reinforcement Learning with Verifiable Rewards (RLVR) without intermediate thinking traces, directly emitting a verdict on code solutions. It specializes in judging and ranking candidate solutions for competitive programming problems, serving as a plug-and-play reward function for code generation policy optimization and automated evaluation.
Loading preview...
Aletheia-Bench/GRPO-Instruct-14B: A Code Verifier
Aletheia-Bench/GRPO-Instruct-14B is a 14.8 billion parameter instruction-tuned code verifier, fine-tuned from Qwen/Qwen2.5-14B-Instruct. It is part of the Aletheia project, which investigates the performance-cost trade-offs in Reinforcement Learning with Verifiable Rewards (RLVR) for code verifiers. This specific model is trained with on-policy learning and negative samples but notably without intermediate thinking traces, directly providing a verdict on code quality.
Key Capabilities
- Code Verification: Judges and ranks candidate solutions for competitive programming problems.
- RLVR Training: Utilizes Reinforcement Learning with Verifiable Rewards (RLVR) for robust performance.
- Efficiency: Designed to operate without generating intermediate thinking traces, offering a compute-efficient strategy.
- Scalability Insights: Developed based on research indicating that eliminating on-policy training at larger model scales can yield comparable performance to full RLVR recipes.
Good For
- RLHF / RLAIF: Serving as a plug-and-play reward function for optimizing code generation policies.
- Automated Evaluation: Functioning as an LLM-as-a-judge for various code-related tasks.
- Research: Studying the impact of thinking traces, on-policy learning, and negative samples in training effective code verifiers.