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

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

ReflectRL-Qwen2.5-Math-7B-GRPO is a model developed by Jinhe Bi and colleagues, trained using the ReflectRL framework. This framework enables learning from Golden Negative Trajectories (GNTs) during on-policy post-training. It utilizes GNTs as reflective context during training, gradually transitioning the policy to direct reasoning for inference, making it suitable for tasks requiring robust learning from failures.

Loading preview...

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

This model is a product of the ReflectRL framework, a lightweight approach designed for learning from Golden Negative Trajectories (GNTs) during on-policy post-training. Developed by Jinhe Bi and a team of researchers, ReflectRL distinguishes itself by not directly imitating failed expert trajectories. Instead, it leverages them as a reflective context during the training phase.

Key Capabilities

  • Learning from Negative Trajectories: Utilizes GNTs to inform the learning process without direct imitation.
  • Reflective-to-Direct Reasoning Transition: Gradually shifts the model's policy from reflective context-based learning to direct reasoning for inference, enhancing robustness.
  • On-Policy Post-Training: Integrates this learning mechanism within an on-policy post-training setup.

Good For

  • Tasks requiring robust learning from failures: Where understanding and avoiding past mistakes is crucial.
  • Research into reinforcement learning and trajectory optimization: Provides a practical implementation of the ReflectRL framework.
  • Developing models that can adapt and improve based on negative examples: Offers a method to incorporate 'what not to do' into the learning process effectively.

Finetunes
2 models