INSAIT-Institute/GRPO-Think-1.5B-4k

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 12, 2026License:cc-by-nc-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

GRPO-Think-1.5B-4k is a 1.5 billion parameter code verifier developed by INSAIT-Institute, fine-tuned from DeepSeek-R1-Distill-Qwen-1.5B. This model is specifically designed to judge and rank candidate solutions for competitive programming problems, utilizing a 4k reasoning-token budget during its GRPO (on-policy, with negative samples and intermediate thinking traces) training. It excels at identifying correct code snippets and is part of the Aletheia research into efficient code verifier deployment.

Loading preview...

Aletheia: Understanding Code Verifiers

This model, GRPO-Think-1.5B-4k, is a specialized code verifier developed by INSAIT-Institute as part of their Aletheia research. It is a 1.5 billion parameter model, fine-tuned from deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B, and trained using the GRPO (on-policy, with negative samples and intermediate thinking traces) algorithm. A key characteristic is its 4k reasoning-token budget, which defines the maximum completion length used during training.

Key Capabilities

  • Code Verification: Judges and ranks candidate solutions for competitive programming problems, identifying the correct one from a set of 2-5 options.
  • RLVR Research: Contributes to the understanding of Reinforcement Learning with Verifiable Rewards (RLVR) for code verifiers, exploring the impact of thinking traces, negative samples, and on-policy training.
  • Scalability Insights: Provides empirical data on how optimal training recipes for code verifiers vary with model scale.

Good For

  • RLHF / RLAIF: Can serve as a plug-and-play reward function for optimizing code generation policies.
  • Automated Evaluation: Useful for LLM-as-a-judge applications in various code-related tasks.
  • Research: Ideal for studying the effects of different training components (thinking traces, on-policy learning, negative samples) on code verifier performance and efficiency.

Finetunes
1 models