Aletheia-Bench/RAFT-7B

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Dec 7, 2025License:cc-by-nc-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

RAFT-7B is a 7.6 billion parameter code verifier model developed by Aletheia-Bench, fine-tuned from DeepSeek-R1-Distill-Qwen-7B. It specializes in judging and ranking candidate code solutions for competitive programming problems, utilizing on-policy rejection sampling fine-tuning with only positive reasoning samples. This model is designed for integration into RLHF/RLAIF pipelines as a reward function for code generation policy optimization and for automated evaluation of code.

Loading preview...

Aletheia-Bench/RAFT-7B: A Code Verifier Model

RAFT-7B is a 7.6 billion parameter code verifier model developed by Aletheia-Bench, derived from deepseek-ai/DeepSeek-R1-Distill-Qwen-7B. It is specifically trained using on-policy rejection sampling fine-tuning, uniquely learning only from positive reasoning samples and excluding negative examples. This approach is part of the broader Aletheia research into Reinforcement Learning with Verifiable Rewards (RLVR) for code verifiers.

Key Capabilities

  • Code Solution Ranking: Given a competitive programming problem and multiple candidate solutions, RAFT-7B can judge and rank these candidates to identify the most correct one.
  • Specialized Training: Utilizes a distinct training recipe (RAFT) focusing on on-policy learning with positive reasoning traces, differing from other verifiers that might incorporate negative samples or offline preference optimization.
  • Execution-Grounded: Trained on the Aletheia-Train dataset, which consists of 50,000 instances of competitive programming problems with execution-verified solutions across Python, C++, and Java.

Good For

  • RLHF / RLAIF: Serving as a plug-and-play reward function to optimize code generation policies.
  • Automated Evaluation: Acting as an LLM-as-a-judge for various code-related tasks, providing objective assessments.
  • Research: Investigating the impact of different training components (thinking traces, on-policy learning, negative samples) on the performance of code verifiers, particularly the efficacy of positive-only learning.

Finetunes
1 models