Nanthasit/sakthai-plus-1.5b
Nanthasit/sakthai-plus-1.5b is a 1.5 billion parameter language model built for agentic tool calling, trained on structured function-calling examples and merged from rsLoRA adapters. It follows the Qwen2.5 chat format and excels at emitting function calls in JSON, optimized for small-footprint CPU and GPU inference. This model is specifically designed for reliable tool selection and JSON argument formatting, making it suitable for applications requiring precise function invocation.
Loading preview...
Overview
Nanthasit/sakthai-plus-1.5b is a 1.5 billion parameter model specifically engineered for agentic tool calling rather than general chat. It was trained on multi-source and structured function-calling examples, leveraging rsLoRA adapters merged into full weights. The model adheres to the Qwen2.5 chat format and is designed to emit function calls in JSON when tool-enabling system prompts are used.
Key Capabilities
- Strong Tool Calling: Achieves 100% success rate in tool call, valid JSON argument, and correct answer rates in verified tests (llama.cpp q4_k_m, 3-trial).
- Efficient Inference: Its small 1.5B parameter size allows for fast inference on both CPUs and consumer GPUs.
- Reliable JSON Output: Demonstrates strong tool selection and consistent JSON argument formatting.
- Cost-Effective Training: Developed using zero paid compute, relying on free-tier T4 credits.
Benchmarks and Limitations
While excelling in tool calling, the model's argument accuracy for complex nested parameters can still be refined. Unverified benchmarks for tasks like Commonsense (WinoGrande 59.6%, HellaSwag 34.0%) and Math (GSM8K 50.9%) are single-trial and indicative, with multi-trial replication planned. It performs best with short- to medium-length tool definitions, as very large schemas may degrade accuracy. Outputs may require parsing with a JSON-tolerant decoder due to potential formatting drift at low temperatures.