metacognitive-behavioral-tuning/Qwen3-0.6B-MBT-S
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
The metacognitive-behavioral-tuning/Qwen3-0.6B-MBT-S is a 0.8 billion parameter language model based on the Qwen3-0.6B architecture. It is fine-tuned using Metacognitive Behavioral Tuning (MBT-S) which involves Supervised Fine-Tuning (SFT) on synthesized metacognitive traces followed by GRPO. This model is specifically optimized for multi-hop question answering tasks, demonstrating performance on benchmarks like HotpotQA, MuSiQue, and 2WikiMultiHopQA.
Loading preview...
Model Overview
The metacognitive-behavioral-tuning/Qwen3-0.6B-MBT-S is a 0.8 billion parameter language model derived from the Qwen/Qwen3-0.6B base. It represents the final checkpoint after a specialized training regimen known as Metacognitive Behavioral Tuning (MBT-S).
Key Capabilities
- Multi-Hop Question Answering: The model is specifically designed and optimized for complex multi-hop question answering, where information must be retrieved and synthesized from multiple sources or steps.
- Metacognitive Behavioral Tuning (MBT-S): Its training involves an initial Supervised Fine-Tuning (SFT) phase on
gpt-oss-120b-synthesized5-phase metacognitive traces, followed by a GRPO (Gradient-based Reward Policy Optimization) stage. This method aims to instill metacognitive reasoning abilities. - Benchmark Performance: Evaluated on multi-hop QA datasets, including HotpotQA (in-domain) and MuSiQue / 2WikiMultiHopQA (out-of-domain), indicating its proficiency in these challenging tasks.
Good For
- Research in Multi-Hop QA: Ideal for researchers exploring advanced reasoning and multi-hop question answering techniques.
- Applications Requiring Complex Information Synthesis: Suitable for use cases where models need to combine information from several steps to form a coherent answer, rather than just retrieving direct facts.
- Exploring Metacognitive Training Methods: Developers interested in models trained with metacognitive approaches will find this model relevant for experimentation and deployment.