Aletheia-Bench/GRPO-Think-14B-4k

TEXT GENERATIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Dec 8, 2025License:cc-by-nc-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Aletheia-Bench/GRPO-Think-14B-4k is a 14.8 billion parameter GRPO-Think code verifier model developed by Aletheia-Bench, fine-tuned from DeepSeek-R1-Distill-Qwen-14B. This model is specifically trained with Reinforcement Learning with Verifiable Rewards (RLVR) using on-policy learning, negative samples, and intermediate thinking traces, with a 4k reasoning-token budget. It excels at judging and ranking candidate solutions for competitive programming problems, making it suitable for automated code evaluation and as a reward function in RLHF/RLAIF for code generation.

Loading preview...

Aletheia-Bench/GRPO-Think-14B-4k: A Code Verifier Model

GRPO-Think-14B-4k is a 14.8 billion parameter code verifier developed by Aletheia-Bench, fine-tuned from deepseek-ai/DeepSeek-R1-Distill-Qwen-14B. It is part of the Aletheia project, which investigates the effectiveness of Reinforcement Learning with Verifiable Rewards (RLVR) for code verifiers. This specific model is trained using the GRPO (Generative Reinforcement Policy Optimization) approach, incorporating on-policy learning, negative samples, and intermediate thinking traces, with a maximum completion length of 4,000 reasoning tokens during training.

Key Capabilities

  • Code Solution Verification: Judges and ranks candidate solutions for competitive programming problems, identifying correct implementations.
  • Thinking Traces: Utilizes intermediate thinking traces during its verification process, enhancing its reasoning capabilities.
  • RLVR Training: Benefits from a robust RLVR pipeline, leveraging both positive and negative samples for improved accuracy.
  • Scalability Insights: Developed as part of research demonstrating that the optimal training recipe for verifiers is scale-dependent, with thinking budget being crucial at larger scales.

Good For

  • RLHF / RLAIF for Code Generation: Can serve as a plug-and-play reward function to optimize code generation policies.
  • Automated Code Evaluation: Ideal for LLM-as-a-judge applications across various code-related tasks.
  • Research on Code Verifiers: Provides a robust platform for studying the impact of thinking traces, on-policy learning, and negative samples in training effective code verifiers.

Finetunes
1 models