Jinhe/ReflectRL-Qwen2.5-1.5B-Instruct-GRPO-ReflectRL

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 4, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Jinhe/ReflectRL-Qwen2.5-1.5B-Instruct-GRPO-ReflectRL is a model developed by Jinhe Bi and colleagues, based on the Qwen2.5-1.5B-Instruct architecture. It implements the ReflectRL framework, which is designed for on-policy post-training by learning from Golden Negative Trajectories (GNTs). This model utilizes GNTs as reflective context during training to improve reasoning, gradually transitioning to direct reasoning for inference.

Loading preview...

ReflectRL-Qwen2.5-1.5B-Instruct-GRPO-ReflectRL Overview

This model integrates the ReflectRL framework with the Qwen2.5-1.5B-Instruct architecture, developed by Jinhe Bi and collaborators. ReflectRL is a lightweight framework focused on enhancing model performance through on-policy post-training, specifically by leveraging Golden Negative Trajectories (GNTs).

Key Capabilities

  • Learning from Negative Trajectories: ReflectRL uniquely uses GNTs not for direct imitation, but as a reflective context during the training process.
  • Reflective-to-Direct Reasoning: The framework facilitates a gradual transition from reflective reasoning during training to direct reasoning during inference, aiming for more robust and efficient decision-making.
  • On-Policy Post-Training: Designed for post-training scenarios where the model learns and adapts based on its own generated trajectories, incorporating insights from failures.

When to Use This Model

This model is particularly relevant for research and applications exploring advanced reinforcement learning techniques, especially those focused on:

  • Improving model robustness by learning from suboptimal or failed experiences.
  • Developing more sophisticated reasoning capabilities in language models.
  • Investigating novel on-policy training methodologies beyond simple imitation learning.

For more technical details, refer to the associated paper and the GitHub repository.