Aletheia-Bench/RAFT-14B
Aletheia-Bench/RAFT-14B is a 14.8 billion parameter code verifier model developed by Aletheia-Bench, fine-tuned from DeepSeek-R1-Distill-Qwen-14B. This RAFT verifier is trained with on-policy rejection sampling using only positive reasoning samples, distinguishing it from models that learn from negative samples or offline preferences. It excels at judging and ranking candidate solutions for competitive programming problems, making it suitable for code generation policy optimization and automated code evaluation.
Loading preview...
Aletheia-Bench/RAFT-14B: A Code Verifier for Competitive Programming
Aletheia-Bench/RAFT-14B is a 14.8 billion parameter code verifier model, fine-tuned from deepseek-ai/DeepSeek-R1-Distill-Qwen-14B. It is part of the Aletheia project, which explores different training recipes for code verifiers using Reinforcement Learning with Verifiable Rewards (RLVR).
Key Capabilities & Differentiators
- RAFT Verifier: This model is specifically a RAFT (Rejection Sampling Fine-Tuning) verifier, trained using on-policy rejection sampling. A key distinction is its learning approach, which utilizes only positive reasoning samples, eschewing negative samples.
- Code Solution Ranking: Given a competitive programming problem and multiple candidate solutions, the model is designed to judge and rank these candidates, identifying the correct one.
- Optimized Training: The Aletheia research highlights that for larger models, the "thinking budget" (intermediate thinking traces) becomes a vital performance factor, while on-policy learning is crucial for smaller verifiers. RAFT-14B incorporates on-policy training with thinking traces.
- Dataset: It was fine-tuned on the
Aletheia-Bench/Aletheia-Traindataset, which comprises 50,000 competitive programming instances with execution-verified correct solutions.
Use Cases
- RLHF / RLAIF: Can serve as a plug-and-play reward function for optimizing code generation policies.
- Automated Evaluation: Suitable for use as an LLM-as-a-judge in various code-related tasks.
- Research: Provides a robust platform for studying the impact of thinking traces, on-policy learning, and negative samples in training effective code verifiers.