Jinhe/ReflectRL-Qwen2.5-Math-7B-DAPO-ReflectRL

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 4, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Jinhe/ReflectRL-Qwen2.5-Math-7B-DAPO-ReflectRL is a model developed by Jinhe Bi and collaborators, based on the Qwen2.5 architecture. This model utilizes the ReflectRL framework, a lightweight method for learning from Golden Negative Trajectories (GNTs) during on-policy post-training. It is specifically designed to improve reasoning by using failed expert trajectories as reflective context during training, transitioning to direct reasoning for inference, making it suitable for tasks requiring robust problem-solving and mathematical capabilities.

Loading preview...

ReflectRL-Qwen2.5-Math-7B-DAPO-ReflectRL Overview

This model, developed by Jinhe Bi and a team of researchers, implements the ReflectRL framework on a Qwen2.5 base. ReflectRL is a novel, lightweight approach designed to enhance model reasoning by leveraging "Golden Negative Trajectories" (GNTs) during the post-training phase. Unlike traditional methods that might directly imitate failed attempts, ReflectRL uses these GNTs as a reflective context during training.

Key Capabilities

  • Learning from Negative Examples: Employs a unique strategy to learn from suboptimal or failed expert trajectories, treating them as reflective insights rather than direct imitation targets.
  • Reflective-to-Direct Reasoning: The training process gradually transitions the model from using reflective context to direct reasoning for inference, aiming for more robust and accurate outputs.
  • On-Policy Post-Training: Integrates this learning mechanism within an on-policy post-training framework, suggesting an iterative refinement process.

Good For

  • Mathematical and Reasoning Tasks: The underlying Qwen2.5-Math base, combined with ReflectRL's focus on learning from complex trajectories, suggests strong performance in tasks requiring logical deduction and problem-solving.
  • Improving Robustness: By analyzing and learning from "negative" examples, the model is likely to develop a more nuanced understanding and avoid common pitfalls.
  • Research in Reinforcement Learning and LLM Alignment: This model serves as a practical implementation of the ReflectRL paper, offering a valuable resource for researchers exploring advanced training techniques for large language models.