INSAIT-Institute/GRPO-Instruct-1.5B
INSAIT-Institute/GRPO-Instruct-1.5B is a 1.5 billion parameter code verifier model developed by INSAIT-Institute. Fine-tuned from Qwen2.5-1.5B-Instruct, it utilizes Reinforcement Learning with Verifiable Rewards (RLVR) to judge and rank candidate solutions for competitive programming problems. This model is specifically trained without intermediate thinking traces, directly emitting a verdict, making it efficient for deploying robust code verifiers in post-training pipelines.
Loading preview...
Aletheia: GRPO-Instruct-1.5B Code Verifier
This model, GRPO-Instruct-1.5B, is a 1.5 billion parameter code verifier developed by INSAIT-Institute as part of the Aletheia project. It is fine-tuned from Qwen/Qwen2.5-1.5B-Instruct using Reinforcement Learning with Verifiable Rewards (RLVR). A key differentiator of this specific verifier is its training approach: it operates without intermediate thinking traces, directly providing a verdict on candidate code solutions. This design choice aims to balance performance and computational cost, particularly for smaller verifiers.
Key Capabilities
- Code Verification: Judges and ranks candidate solutions for competitive programming problems.
- RLVR Training: Utilizes on-policy learning with negative samples for robust verification.
- Efficiency: Designed to directly emit a verdict without generating intermediate thought processes, offering a compute-efficient strategy at lower budgets.
- Scalability Insights: Part of a broader research effort to understand how different RLVR training recipes (thinking traces, negative samples, on-policy training) impact verifier performance across various model scales.
Good For
- RLHF / RLAIF: Serving as a plug-and-play reward function for optimizing code generation policies.
- Automated Evaluation: Acting as an LLM-as-a-judge for diverse code-related tasks.
- Research: Studying the impact of specific training components (like the absence of thinking traces) on code verifier efficacy and efficiency.