SakThai Context Paper

Building Efficient Tool-Calling Language Models for Agentic Workflows

By Nanthasit Burankum (Beer) — House of Sak

Abstract

This paper presents the SakThai Context family — fine-tuned Qwen2.5 variants for tool-calling, multi-turn context, and instruction following. Three sizes (0.5B, 1.5B, 7B) achieve 100% pass rates on custom evaluation suites. Includes a 128K context extension via YaRN.

The models are trained on the SakThai combined dataset series (v1–v5), covering 25+ tool schemas across diverse domains. Training uses QLoRA for efficiency, with adapters merged into full-weight checkpoints for deployment.

Key Findings

✅ Small models can tool-call effectively — 0.5B achieves 100% on Eval Suite
✅ QLoRA is sufficient — no full fine-tuning needed for tool-calling
✅ YaRN extends context without retraining — 128K from 32K with config only
✅ T4 GPU compatible — 7B model uses only 5.56 GB VRAM

Read the Full Paper

File	Description
PAPER.md	Full whitepaper — 10 sections, ~14K words

Citation

@misc{sakthai-context-paper,
  author = {Nanthasit Burankum},
  title = {SakThai Context: Building Efficient Tool-Calling Language Models for Agentic Workflows},
  year = {2026},
  publisher = {Hugging Face},
  journal = {House of Sak},
  howpublished = {https://huggingface.co/Nanthasit/sakthai-context-paper}
}

Related Models

Model	Description
0.5B merged	Lightweight tool-calling
1.5B merged	Most popular
7B merged	Best performance
7B 128K	Long context