yapeichang/Qwen2.5-7B-RM8B

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 27, 2025License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The yapeichang/Qwen2.5-7B-RM8B model is a 7.6 billion parameter language model based on the Qwen2.5-7B architecture, developed by Yapei Chang and collaborators. It is specifically fine-tuned using GRPO with the Skywork-RM-8B reward model, focusing on general instruction following. This model demonstrates performance comparable to RM-guided GRPO across various instruction-following benchmarks, producing factually grounded outputs.

Loading preview...

Overview

This model, yapeichang/Qwen2.5-7B-RM8B, is a 7.6 billion parameter language model built upon the Qwen2.5-7B architecture. It was developed by Yapei Chang and a team of researchers, as detailed in their paper "BLEUBERI: BLEU is a surprisingly effective reward for instruction following." The model is fine-tuned using a novel approach called GRPO (Generalized Reward Policy Optimization) where BLEU, a simple n-gram matching metric, is directly used as a reward signal.

Key Capabilities

  • Enhanced Instruction Following: The model is specifically optimized for general instruction following tasks, extending beyond easily verifiable domains like math and code.
  • Factually Grounded Outputs: It is designed to produce more factually grounded responses, with human evaluations rating its outputs on par with systems trained using traditional reward models.
  • GRPO Training with BLEU: Utilizes a unique training methodology where BLEU, paired with high-quality references from strong LLMs, serves as an effective reward function.

Good For

  • Applications requiring robust general instruction following capabilities.
  • Generating responses that are both coherent and factually consistent.
  • Research and development in reward modeling and reinforcement learning from human feedback (RLHF) alternatives.