Aletheia-Bench/GRPO-Instruct-1.5B
Aletheia-Bench/GRPO-Instruct-1.5B is a 1.5 billion parameter instruction-tuned code verifier developed by Aletheia-Bench, fine-tuned from Qwen2.5-1.5B-Instruct. This model is trained using Reinforcement Learning with Verifiable Rewards (RLVR) without intermediate thinking traces, directly emitting a verdict. It excels at 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...
Model Overview
Aletheia-Bench/GRPO-Instruct-1.5B is a 1.5 billion parameter instruction-tuned code verifier, fine-tuned from Qwen/Qwen2.5-1.5B-Instruct. It is part of the Aletheia project, which explores the effectiveness of various RLVR (Reinforcement Learning with Verifiable Rewards) training recipes for code verifiers. This specific model is a GRPO-Instruct verifier, meaning it was trained with on-policy learning and negative samples but without intermediate thinking traces, directly providing a verdict on code solutions.
Key Capabilities
- Code Verification: Judges and ranks candidate solutions for competitive programming problems, identifying the correct one among several options.
- RLVR Training: Utilizes Reinforcement Learning with Verifiable Rewards, specifically the GRPO-Instruct approach, which is optimized for efficiency by omitting intermediate thinking traces.
- Scalability Insights: Part of a research effort to understand how different RLVR components (thinking traces, negative samples, on-policy training) impact performance across various model scales.
Good For
- RLHF / RLAIF: Can be used as a plug-and-play reward function to optimize code generation policies.
- Automated Evaluation: Suitable for LLM-as-a-judge applications in various code-related tasks.
- Research: Ideal for studying the impact of different training methodologies on code verifier performance, particularly the trade-offs between training cost and accuracy.