metacognitive-behavioral-tuning/Qwen3-0.6B-GRPO

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 3, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Qwen3-0.6B-GRPO is a 0.8 billion parameter causal language model developed by metacognitive-behavioral-tuning, based on the Qwen3 architecture. This model is specifically fine-tuned using the GRPO method for multi-hop question answering tasks, demonstrating specialized performance on datasets like HotpotQA. It features a context length of 32768 tokens and is optimized for complex reasoning over multiple pieces of information.

Loading preview...

Qwen3-0.6B-GRPO: Multi-Hop Question Answering Model

Qwen3-0.6B-GRPO is a specialized 0.8 billion parameter language model developed by metacognitive-behavioral-tuning, built upon the Qwen/Qwen3-0.6B base architecture. Its primary distinction lies in its training methodology, utilizing the GRPO (Gradient-based Reward Policy Optimization) technique. Unlike standard fine-tuning, this model was trained directly with GRPO after an initial Supervised Fine-Tuning (SFT) stage, specifically targeting enhanced performance in multi-hop question answering.

Key Capabilities and Training:

  • Multi-Hop Question Answering: Optimized for tasks requiring reasoning across multiple pieces of information to formulate an answer.
  • GRPO Training: Employs a unique GRPO training method, detailed in the paper "Metacognitive Behavioral Tuning of Large Language Models for Multi-Hop Question Answering."
  • Base Model: Leverages the robust Qwen/Qwen3-0.6B as its foundation.
  • Training Process: Underwent an SFT phase (LR 1e-4, BS 128, on HotpotQA) followed by the GRPO stage.
  • Context Length: Supports a substantial context window of 32768 tokens.

Benchmarking and Use Cases:

  • Evaluated on: HotpotQA (in-distribution) and MuSiQue / 2WikiMultiHopQA (out-of-distribution) datasets.
  • Ideal for: Applications requiring advanced reasoning and information synthesis from multiple sources, particularly in question-answering systems where direct answers are not immediately apparent.