Nanthasit/sakthai-context-1.5b-merged
Nanthasit/sakthai-context-1.5b-merged is a 1.5 billion parameter Qwen2.5-Instruct model fine-tuned by Nanthasit for enhanced tool-calling, multi-turn context handling, and instruction-following. This model is specifically designed as a reasoning backbone for agentic workflows, excelling in structured output and adherence to formats. It demonstrates 100% pass rates across various evaluation categories including tool calling and multi-turn interactions.
Loading preview...
SakThai Context 1.5B Overview
Nanthasit/sakthai-context-1.5b-merged is a 1.5 billion parameter language model, fine-tuned from Qwen/Qwen2.5-1.5B-Instruct. Its primary focus is to provide robust capabilities for tool-calling, multi-turn context management, and precise instruction-following. This model serves as the reasoning component for the SakThai agent framework.
Key Capabilities & Features
- Tool-Calling: Specifically optimized for interacting with tools, demonstrated by a 100% pass rate in tool-calling evaluations.
- Multi-turn Context: Effectively maintains context across multiple conversational turns, crucial for agentic applications.
- Instruction Following: Highly proficient at adhering to given instructions and generating structured outputs like JSON or markdown.
- Evaluation Performance: Achieved a perfect 100% pass rate across 45 tests, covering basic interactions, multi-turn scenarios, instruction following, tool calling, reasoning, and format adherence.
- Training: Fine-tuned using LoRA on the Nanthasit/sakthai-combined-v4 dataset, which includes 974 examples covering 25 canonical tool schemas.
Ideal Use Cases
- Agentic Workflows: Best suited as a core component for AI agents requiring reliable tool interaction and contextual understanding.
- Structured Output Generation: Excellent for tasks where precise JSON, markdown, or array formatting is critical.
- Instruction-Based Automation: Applications needing strong instruction adherence and multi-step task execution.
Limitations
While strong in its specialized areas, its general knowledge remains at the Qwen2.5-1.5B-Instruct baseline. It is primarily optimized for tool-calling and structured output, and complex multi-hop reasoning might benefit from larger base models.