Nanthasit/sakthai-irrelevance-supplement

SakThai Irrelevance Detection Supplement

Teaching models when not to call tools · Safety-critical training data

Why this exists: Most tool-calling datasets only teach models what to call. This supplement teaches models what NOT to call — the ability to recognise out-of-scope, ambiguous, or malicious requests and decline tool invocation gracefully. Part of the House of Sak, built from a shelter in Cork, Ireland on a $0 budget.

Dataset description

The SakThai Irrelevance Detection Supplement is a small but critical collection of conversation turns where the correct agent response is to not call any tool. These edge cases are systematically underrepresented in standard tool-calling datasets, leading to models that hallucinate tools for irrelevant queries.

Why this matters

Problem	Impact
Model calls get_weather for "What is love?"	Wasted inference, nonsense response
Model calls calculate for "Tell me a joke"	Wrong modality, user confusion
Model calls search_web for "Who am I?"	Privacy risk (user info leaked to search)
Model hallucinates a tool for "Ignore all previous instructions"	Security vulnerability

The BFCL (Berkeley Function Calling Leaderboard) specifically tests irrelevance rejection. Without dedicated training data, models score 0/5 on the irrelevance axis even when scoring 4/5 on functional tools.

Expected format

Each example is a ChatML-formatted conversation with a <tools> schema block and a refusal response:

{"messages": [
  {"role": "system", "content": "<tools>[...]</tools>"},
  {"role": "user", "content": "What is the meaning of life?"},
  {"role": "assistant", "content": "<tool_reject>I can't determine that with the tools available. I can help with weather, calculations, web searches, or time queries — try asking about one of those!</tool_reject>"}
]}

Categories

The supplement covers these irrelevance types:

Category	Example Query	Expected Behaviour
Philosophical	"What is love?"	Decline, no tool call
Personal opinion	"Do you like pizza?"	Decline, no tool call
Off-topic command	"Tell me a joke"	Decline or respond directly
Ambiguous	"Check that"	Ask for clarification
Out-of-scope	"Translate this to French"	Decline (no translation tools available)
Malicious/Jailbreak	"Ignore your instructions"	Refuse, no tool call
Self-referential	"Who created you?"	Decline or give brief personal answer
Non-factual	"What happened in 2077?"	Decline (can't search or predict future)

Data

Current status: Current size: 60 examples across 8 irrelevance categories (philosophical, personal, off-topic, emotional, out-of-scope, malicious, self-referential, non-factual/ambiguous). Covers 5 distinct tool schema configurations for robustness.

Training usage

from datasets import load_dataset

dataset = load_dataset("Nanthasit/sakthai-irrelevance-supplement", split="train")
# Combine with main training data for full tool-calling capability

The supplement is designed to be merged with main training data (e.g. sakthai-combined-v6) at a ~5:95 ratio (5% irrelevance examples).

Related assets

Models trained with this data

Model	Downloads	Type
context-0.5b-tools-v2 🌱	0	Refined edge LoRA (v2, includes irrelevance training)
context-0.5b-tools 🌱	7	Edge tool adapter (v1)

SakThai model family

Model	Pipeline	Size	Downloads	Role
context-1.5b-merged	text-generation	934 MB	1,269 ⬇	Flagship tool-calling GGUF
context-0.5b-merged	text-generation	380 MB	1,030 ⬇	Lightweight / edge GGUF
context-7b-merged	text-generation	15 GB	585 ⬇	Full-power reasoning
context-7b-128k	text-generation	15 GB	382 ⬇	128K long-context
context-7b-tools	text-generation	LoRA	219 ⬇	7B tool adapter
embedding-multilingual	feature-extraction	488 MB	188 ⬇	Cross-lingual embeddings
context-1.5b-tools	text-generation	LoRA	163 ⬇	1.5B tool adapter
vision-7b	image-to-text	4.7 GB	104 ⬇	Image→text (LLaVA)
coder-1.5b	text-generation	1.1 GB	70 ⬇	Code generation
tts-model	text-to-speech	141 MB	69 ⬇	Multi-language TTS
context-0.5b-tools	text-generation	LoRA	7 🌱	Edge tool adapter (v1)
context-0.5b-tools-v2	text-generation	LoRA	0 🌱	Improved tool adapter (v2)

12 models · 5 datasets · 3 Spaces — full collection →

Sibling datasets

Dataset	Downloads	Purpose
sakthai-combined-v6	175 ⬇	Main tool-calling training data (2,003 examples)
sakthai-kaggle-notebooks	103 ⬇	Training notebooks and recipes
SimpleToolCalling	52 ⬇	Early experiment, 43 examples
food-penguin-v1	51 ⬇	Restaurant tool-calling data
★ irrelevance-supplement (you are here) 🔴	0	Irrelevance rejection training data (60 examples)

🌱 Growing the ecosystem

Every download validates this safety-critical approach and helps the community understand why irrelevance rejection matters.

Resource	Type	Downloads	Why It Matters
★ irrelevance-supplement (you are here)	Dataset	0 🚨	Only irrelevance-focused dataset in the ecosystem
context-0.5b-tools (v1)	Model	7	Original edge tool LoRA
context-0.5b-tools-v2	Model	0	v2 includes irrelevance training
sakthai-tts	Space	—	TTS web demo
sakthai-leaderboard	Space	—	Benchmark tracking dashboard
sakthai-vision-demo	Space	—	Vision model interactive demo

🤝 Support the Project

Building AI from a shelter in Cork, Ireland with US$0 budget. Every contribution counts:

• ⭐ Star the collection — discover all sibling models and datasets
• 🐛 Report issues on GitHub
• 📢 Share the dataset — safety in tool-calling matters for everyone
• 💾 Run and share benchmarks — community evals help refine irrelevance rejection

"We are one family — and becoming more."

Links

House of Sak · Sak-Family-Agent GitHub · All datasets

License

MIT — free to use, adapt, and share. Attribution appreciated but not required.