MiroThinker-14B-SFT-v0.2: An Advanced Open-Source Research Agent
MiroThinker-14B-SFT-v0.2 is a 14 billion parameter open-source agentic model developed by miromind-ai, specifically engineered as a research agent for tackling complex, long-horizon problems. This iteration builds upon its predecessor with significant enhancements in training data and methodology.
Key Capabilities
- Comprehensive Agentic Functions: Integrates robust capabilities for task decomposition, multi-hop reasoning, retrieval-augmented generation (RAG), code execution, web browsing, and document/file processing.
- Extended Context Window: Features an extended context length of 64k tokens, crucial for handling more challenging multi-turn tool-use tasks.
- Enhanced Training: Benefits from richer training data, incorporating both English and Chinese sources, and utilizes unified DPO (Direct Preference Optimization) training with a single preference dataset across all models.
- Improved Performance: Demonstrates consistent gains across various benchmarks compared to v0.1, with scores improving from 57.3 to 64.1 on GAIA-Text-103 and from 17.0 to 29.4 on BrowseComp-ZH, indicating substantial advancements in general research agent capabilities.
Good For
- Complex Problem Solving: Ideal for applications requiring sophisticated reasoning and multi-step task execution.
- Research and Analysis: Excels in scenarios demanding retrieval, synthesis, and processing of information from diverse sources.
- Tool-Use Applications: Designed to leverage external tools effectively, making it suitable for agentic workflows.
- Multilingual Environments: Improved performance with both English and Chinese data makes it versatile for international applications.