Jinhe/ReflectRL-Llama-3.1-8B-Instruct-GRPO

TEXT GENERATIONPricing:Input $0.2 / Cached $0.028 / Output $0.32Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 4, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Jinhe/ReflectRL-Llama-3.1-8B-Instruct-GRPO is an 8 billion parameter instruction-tuned model based on Llama-3.1, developed by Jinhe Bi and collaborators. It utilizes the ReflectRL framework for on-policy post-training, learning from Golden Negative Trajectories (GNTs) by using them as reflective context. This approach allows the model to transition from reflective reasoning during training to direct reasoning for inference, making it suitable for tasks requiring robust learning from failed attempts.

Loading preview...

ReflectRL-Llama-3.1-8B-Instruct-GRPO Overview

This model, developed by Jinhe Bi and collaborators, is an 8 billion parameter instruction-tuned variant of Llama-3.1. It incorporates the ReflectRL framework, a lightweight method designed for learning from Golden Negative Trajectories (GNTs) during on-policy post-training. Unlike traditional approaches that might directly imitate failed expert trajectories, ReflectRL leverages these GNTs as a reflective context during the training phase.

Key Capabilities & Differentiators

  • Learning from Golden Negative Trajectories (GNTs): ReflectRL specifically utilizes information from failed trajectories to improve model performance, rather than just successful ones.
  • Reflective-to-Direct Reasoning: The framework enables the model to use reflective reasoning during training, gradually transitioning to a more direct reasoning approach for inference, optimizing for efficiency and performance.
  • Lightweight Framework: ReflectRL is described as a lightweight approach for post-training, suggesting potential efficiency benefits.

Good For

  • Tasks requiring robust learning from failures: Ideal for scenarios where understanding and avoiding past mistakes is crucial for improved performance.
  • Research into reinforcement learning from human feedback (RLHF) extensions: Provides a practical implementation of learning from negative examples.
  • Applications benefiting from nuanced reasoning: The reflective-to-direct reasoning transition could be beneficial for complex problem-solving where initial reflection on failures leads to better direct solutions.

Finetunes
2 models