metacognitive-behavioral-tuning/Qwen3-1.7B-GRPO
metacognitive-behavioral-tuning/Qwen3-1.7B-GRPO is a 2 billion parameter language model based on the Qwen3 architecture, developed by metacognitive-behavioral-tuning. This model is specifically fine-tuned using the GRPO method for multi-hop question answering tasks. It excels at complex information retrieval and reasoning, particularly on datasets like HotpotQA, MuSiQue, and 2WikiMultiHopQA. With a context length of 32768 tokens, it is designed for deep contextual understanding in question-answering scenarios.
Loading preview...
Model Overview
metacognitive-behavioral-tuning/Qwen3-1.7B-GRPO is a 2 billion parameter language model built upon the Qwen/Qwen3-1.7B base model. Its core innovation lies in its training methodology: it was directly trained using the GRPO (Gradient-based Reward Policy Optimization) method, following an initial Supervised Fine-Tuning (SFT) stage. This approach is detailed in the paper Metacognitive Behavioral Tuning of Large Language Models for Multi-Hop Question Answering.
Key Capabilities
- Multi-Hop Question Answering: Specifically optimized for tasks requiring reasoning across multiple pieces of information to formulate an answer.
- GRPO Training: Utilizes a unique training paradigm that enhances its ability to handle complex question-answering scenarios.
- Context Length: Supports a substantial context window of 32768 tokens, enabling it to process and synthesize information from lengthy inputs.
Benchmarks and Performance
The model's performance is evaluated on several multi-hop question answering benchmarks:
- HotpotQA: Used for in-domain (ID) evaluation, indicating its proficiency on similar datasets.
- MuSiQue / 2WikiMultiHopQA: Used for out-of-domain (OOD) evaluation, demonstrating its generalization capabilities to new, unseen multi-hop question answering tasks.
Good For
- Research in Multi-Hop QA: Ideal for researchers exploring advanced reasoning and question-answering techniques.
- Applications requiring complex information synthesis: Suitable for use cases where answers are not directly stated but require combining information from multiple sources.