INSAIT-Institute/RAFT-14B

TEXT GENERATIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 12, 2026License:cc-by-nc-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

RAFT-14B is a 14.8 billion parameter code verifier model developed by INSAIT-Institute, fine-tuned from DeepSeek-R1-Distill-Qwen-14B. It is specifically trained using on-policy rejection sampling fine-tuning, learning exclusively from positive reasoning samples. This model 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...

Overview

RAFT-14B is a 14.8 billion parameter code verifier developed by INSAIT-Institute, part of the Aletheia project. It is fine-tuned from deepseek-ai/DeepSeek-R1-Distill-Qwen-14B using a specific training recipe: on-policy rejection sampling fine-tuning that leverages only positive reasoning samples (excluding negative examples). This approach aims to efficiently train robust code verifiers, addressing the high costs typically associated with full Reinforcement Learning with Verifiable Rewards (RLVR) pipelines.

Key Capabilities

  • Code Solution Verification: Judges and ranks candidate solutions for competitive programming problems, given the problem statement and multiple code submissions.
  • Specialized Training: Utilizes a unique RAFT (Rejection Sampling Fine-Tuning) algorithm, focusing on positive reasoning samples for efficiency and performance.
  • Foundation for RL: Designed to serve as a plug-and-play reward function for optimizing code generation policies through RLHF/RLAIF.
  • Automated Evaluation: Can function as an LLM-as-a-judge for various code-related tasks, providing objective assessments.

When to Use This Model

  • RLHF/RLAIF for Code: Integrate as a reward function to guide and improve code generation models.
  • Automated Code Assessment: For tasks requiring automated evaluation and ranking of code solutions.
  • Research on Code Verifiers: Ideal for studying the impact of different training components (like thinking traces, on-policy learning, and negative samples) on code verifier performance and large language model performance.

Finetunes
1 models