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

Overview

Qwen3-1.7B-MBT-R is a 2 billion parameter language model derived from the Qwen/Qwen3-1.7B base model, developed by metacognitive-behavioral-tuning. It is specifically engineered for multi-hop question answering through a novel training methodology called Metacognitive Behavioral Tuning (MBT-R).

Key Capabilities

• Enhanced Multi-Hop Reasoning: The model's core strength lies in its ability to process and answer complex questions requiring information synthesis from multiple sources or steps.
• Metacognitive Behavioral Tuning (MBT-R): This method involves rewriting the student model's own reasoning traces into a 5-phase structure for Supervised Fine-Tuning (SFT), followed by Guided Reinforcement Learning with Policy Optimization (GRPO).
• Robust Training: The model underwent SFT with a learning rate of 1e-4 and a batch size of 128 on the HotpotQA dataset, subsequently refined with GRPO.

Benchmarks & Performance

Qwen3-1.7B-MBT-R demonstrates strong performance on multi-hop question answering benchmarks, including:

• HotpotQA: Evaluated for in-domain (ID) performance.
• MuSiQue / 2WikiMultiHopQA: Assessed for out-of-domain (OOD) generalization capabilities.

When to Use

This model is particularly well-suited for applications requiring advanced reasoning over multiple pieces of information to answer complex questions, such as:

• Complex Question Answering Systems
• Information Retrieval with Reasoning
• Knowledge Graph Querying

Its specialized tuning makes it a strong candidate for tasks where understanding and synthesizing information across several steps is critical.