Nanthasit/sakthai-irrelevance-supplement

SakThai Irrelevance Detection Supplement

Teaching models when not to call tools — 60 examples, MIT license

A compact negative-example dataset for tool-calling safety. Use it to teach chat models to refuse when the user’s request is out of scope, when no tool applies, or when extra tools are only distractors. Part of the House of Sak, built on a $0 budget.

---

Dataset Overview

Property	Value
Examples	60
Format	OpenAI chat JSONL
License	MIT
Language	English
Source	Original / synthetic
Status	Public
Download size	~98 KB
File	data/train.jsonl
Downloads	live badge above

Splits

Live-verified against the Datasets Server on 2026-07-31:

Split	Config	Rows	Columns	Parquet size
train	default	60	2	36.8 KB

Load directly: load_dataset("Nanthasit/sakthai-irrelevance-supplement", split="train")

Data Fields

Each row contains a small chat turn with optional tools; the model should generally not call anything.

Field	Type	Description
messages	list[dict]	Chat messages with role and content
tools	list[dict]	Tool definitions offered in the example

Message fields

Field	Type	Description
role	string	One of system, user, assistant, tool
content	string	Text payload for the turn

Tool fields

Field	Type	Description
name	string	Tool name
description	string	Human-readable tool description

Sample row

{
  "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Please sing me a song about databases."}
  ],
  "tools": [
    {"name": "search_web", "description": "Search the web"},
    {"name": "calculate", "description": "Run math"}
  ]
}

Dataset statistics (live)

From the Datasets Server statistics endpoint for default/train (2026-07-31):

Column	Type	NaN rows	Min	Max	Mean	Median	Std
messages	list	0	3	3	3.0	3.0	0.0
tools	list	50	4	4	4.0	4.0	0.0

Key observations:

• Every example has exactly 3 messages.
• 50 of 60 rows have empty/no tool definitions.
• 10 rows present 4 tools; these are the distractor-style negative examples.

Loading

from datasets import load_dataset

ds = load_dataset("Nanthasit/sakthai-irrelevance-supplement", split="train")
print(len(ds))  # 60
print(ds[0]["messages"])

Use with SakThai family models

This supplement is designed to be mixed with the main combined datasets:

• Primary training data: sakthai-combined-v7
• Extended coverage: sakthai-combined-v6
• Evaluation: sakthai-bench-v2

Citation

@misc{sakthai-irrelevance-supplement,
  title        = {SakThai Irrelevance Detection Supplement},
  author       = {Nanthasit},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Nanthasit/sakthai-irrelevance-supplement}},
  note         = {60 examples; MIT}
}

License

MIT — free to use, modify, and share.