INSAIT-Institute/GRPO-Instruct-14B

TEXT GENERATIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 12, 2026License:cc-by-nc-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

INSAIT-Institute/GRPO-Instruct-14B is a 14.8 billion parameter code verifier model developed by INSAIT-Institute, fine-tuned from Qwen2.5-14B-Instruct. It is trained using Reinforcement Learning with Verifiable Rewards (RLVR), specifically on-policy with negative samples, but without intermediate thinking traces, directly emitting a verdict. This model 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...

INSAIT-Institute/GRPO-Instruct-14B: A Code Verifier Model

This model, GRPO-Instruct-14B, is a 14.8 billion parameter code verifier developed by INSAIT-Institute. It is fine-tuned from Qwen/Qwen2.5-14B-Instruct and trained using Reinforcement Learning with Verifiable Rewards (RLVR). A key differentiator of this specific model is its training approach: it utilizes on-policy learning with negative samples but without intermediate thinking traces, directly providing a verdict on code solutions. This design choice aims to optimize the performance-cost trade-off in RLVR for code verification.

Key Capabilities

  • Code Solution Verification: Judges and ranks candidate solutions for competitive programming problems, given a problem statement and multiple code options.
  • RLVR Training: Leverages on-policy training and negative samples for robust verification, based on the Aletheia research framework.
  • Efficiency: Designed to operate without generating intermediate thinking traces, offering a compute-efficient strategy at lower budgets while maintaining strong verifier accuracy.

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 various code-related tasks.
  • Research: Studying the impact of different RLVR components (like thinking traces, on-policy learning, and negative samples) on code verifier performance, particularly at larger model scales.

Finetunes
1 models