Aletheia-Bench/GRPO-Think-14B-16k

TEXT GENERATIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Nov 3, 2025License:cc-by-nc-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Aletheia-Bench/GRPO-Think-14B-16k is a 14.8 billion parameter GRPO-Think code verifier developed by Aletheia-Bench, fine-tuned from DeepSeek-R1-Distill-Qwen-14B. This model is trained with Reinforcement Learning with Verifiable Rewards (RLVR) using on-policy learning, negative samples, and intermediate thinking traces, with a 16k reasoning-token budget. It excels at judging and ranking candidate solutions for competitive programming problems, providing a robust reward function for code generation policy optimization.

Loading preview...

Aletheia-Bench/GRPO-Think-14B-16k: A Code Verifier for RLVR

This model, GRPO-Think-14B-16k, is a 14.8 billion parameter code verifier developed by Aletheia-Bench, fine-tuned from deepseek-ai/DeepSeek-R1-Distill-Qwen-14B. It is specifically designed as a "thinking" verifier, trained using the GRPO (Reinforcement Learning with Verifiable Rewards) approach, which incorporates on-policy learning, negative samples, and intermediate thinking traces. A key feature is its 16k reasoning-token budget, defining the maximum completion length during training.

Key Capabilities

  • Code Solution Verification: Judges and ranks candidate solutions for competitive programming problems, given a problem statement and multiple potential solutions.
  • RLVR Integration: Trained with a comprehensive RLVR pipeline, making it suitable as a plug-and-play reward function for optimizing code generation policies.
  • Robustness: Leverages negative samples and on-policy training to enhance selection accuracy and stabilize training, particularly at larger model scales.
  • Thinking Traces: Utilizes intermediate thinking traces during training, which are crucial for performance in larger verifiers.

Good For

  • RLHF / RLAIF: Serving as a reward function for fine-tuning code generation models.
  • Automated Evaluation: Acting as an LLM-as-a-judge for various code-related tasks.
  • Research: Studying the impact of different training components (thinking traces, negative samples, on-policy learning) on code verifier performance and efficiency. The model is part of the Aletheia testbed, which facilitates contamination-free analysis of code verifier training recipes.

Finetunes
1 models