Aletheia-Bench/GRPO-Think-1.5B-8k

Hugging Face
TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Dec 5, 2025License:cc-by-nc-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Warm

Aletheia-Bench/GRPO-Think-1.5B-8k is a 1.5 billion parameter code verifier model developed by Aletheia-Bench, fine-tuned from DeepSeek-R1-Distill-Qwen-1.5B. This GRPO-Think verifier is trained with Reinforcement Learning with Verifiable Rewards (RLVR), incorporating on-policy learning, negative samples, and intermediate thinking traces, with an 8k reasoning-token budget. It specializes in judging and ranking candidate solutions for competitive programming problems, providing a plug-and-play reward function for code generation policy optimization and automated evaluation.

Loading preview...

Aletheia-Bench/GRPO-Think-1.5B-8k: A Code Verifier Model

GRPO-Think-1.5B-8k is a 1.5 billion parameter code verifier model developed by Aletheia-Bench, fine-tuned from deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B. It is a GRPO-Think verifier, meaning it was trained using Reinforcement Learning with Verifiable Rewards (RLVR) with on-policy learning, negative samples, and intermediate thinking traces. A key characteristic is its 8k reasoning-token budget (maximum completion length) used during training.

Key Capabilities

  • Code Verification: Judges and ranks candidate solutions for competitive programming problems, given a problem statement and multiple solutions.
  • RLVR Training: Incorporates a robust training methodology that includes on-policy learning, negative samples, and intermediate thinking traces, which are crucial for performance, especially at smaller scales.
  • Scalability Insights: Part of the broader Aletheia research, which investigates how optimal training recipes for code verifiers vary with model scale, highlighting the importance of on-policy learning for smaller models like this one.

Good For

  • RLHF / RLAIF: Serving as a plug-and-play reward function to optimize code generation policies.
  • Automated Evaluation: Utilizing an LLM-as-a-judge approach for various code-related tasks.
  • Research: Studying the impact of thinking traces, on-policy learning, and negative samples in the development of effective code verifiers.

Finetunes
1 models