metacognitive-behavioral-tuning/Qwen3-1.7B-MBT-S

TEXT GENERATIONConcurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 3, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

metacognitive-behavioral-tuning/Qwen3-1.7B-MBT-S is a 2 billion parameter Qwen3-based language model developed by metacognitive-behavioral-tuning. This model is fine-tuned using Metacognitive Behavioral Tuning (MBT-S) with Supervised Fine-Tuning (SFT) and Guided Reinforcement Learning from Policy Optimization (GRPO) on metacognitive traces. It specializes in multi-hop question answering, demonstrating performance on datasets like HotpotQA, MuSiQue, and 2WikiMultiHopQA, and supports a 32768 token context length.

Loading preview...

Overview

metacognitive-behavioral-tuning/Qwen3-1.7B-MBT-S is a 2 billion parameter model built upon the Qwen/Qwen3-1.7B base architecture. It implements a novel Metacognitive Behavioral Tuning (MBT-S) approach, which involves an initial Supervised Fine-Tuning (SFT) phase on synthesized 5-phase metacognitive traces from gpt-oss-120b-synthesized, followed by Guided Reinforcement Learning from Policy Optimization (GRPO).

Key Capabilities

  • Multi-Hop Question Answering: Specifically designed and optimized for complex multi-hop question answering tasks, as detailed in the associated paper "Metacognitive Behavioral Tuning of Large Language Models for Multi-Hop Question Answering."
  • Metacognitive Reasoning: Leverages metacognitive traces during training to enhance its reasoning capabilities, particularly for tasks requiring multiple steps of inference.
  • Benchmarked Performance: Evaluated on standard multi-hop QA datasets including HotpotQA (in-distribution) and MuSiQue / 2WikiMultiHopQA (out-of-distribution).

Good For

  • Research in Advanced Reasoning: Ideal for researchers exploring metacognitive learning, multi-hop reasoning, and complex question answering in LLMs.
  • Applications Requiring Multi-Step Inference: Suitable for use cases where models need to synthesize information from multiple sources or steps to arrive at an answer.
  • Building on Qwen3 Architecture: Provides a specialized fine-tuned version of the Qwen3-1.7B base model for specific reasoning challenges.

Finetunes
1 models