Aletheia-Bench/GRPO-Instruct-7B

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Nov 26, 2025License:cc-by-nc-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Aletheia-Bench/GRPO-Instruct-7B is a 7.6 billion parameter code verifier model fine-tuned from Qwen2.5-7B-Instruct by Aletheia-Bench. It is trained using Reinforcement Learning with Verifiable Rewards (RLVR) on the Aletheia-Train dataset, specifically designed to judge and rank candidate solutions for competitive programming problems. This model is distinguished by its RLVR training without intermediate thinking traces, directly emitting a verdict, making it efficient for automated code evaluation and reward function generation in RLHF/RLAIF pipelines.

Loading preview...

Aletheia-Bench/GRPO-Instruct-7B: A Code Verifier Model

GRPO-Instruct-7B is a 7.6 billion parameter code verifier developed by Aletheia-Bench, fine-tuned from Qwen/Qwen2.5-7B-Instruct. This model is part of the Aletheia project, which investigates the effectiveness of Reinforcement Learning with Verifiable Rewards (RLVR) for code verifiers.

Key Capabilities and Differentiators

  • Code Verification: Designed to judge and rank candidate solutions for competitive programming problems, given the problem statement and multiple code snippets.
  • RLVR Training without Thinking Traces: Unlike some other verifiers, GRPO-Instruct-7B is trained with RLVR (on-policy, with negative samples) but without intermediate thinking traces, directly providing a verdict. This approach offers a compute-efficient strategy at lower budgets, balancing training cost and verifier accuracy.
  • Foundation Model: It serves as a practical example from the Aletheia research, which explores optimal training recipes for code verifiers across different scales and scenarios.
  • Dataset: Trained on the Aletheia-Bench/Aletheia-Train dataset, comprising 50,000 competitive programming instances with execution-verified solutions.

Intended Use Cases

  • RLHF / RLAIF: Can be used as a plug-and-play reward function for optimizing code generation policies.
  • Automated Evaluation: Suitable for LLM-as-a-judge applications in various code-related tasks.
  • Research: Valuable for studying the impact of thinking traces, on-policy learning, and negative samples in the development of robust code verifiers.

Finetunes
1 models