Nanthasit/sakthai-combined-v6

SakThai Combined Dataset v6 🎯

Tool-calling, multi-turn, safety, and irrelevance — the primary training set behind the SakThai model family

The primary training set behind the SakThai model family — tool-calling, multi-turn, irrelevance, and safety examples in OpenAI chat JSONL format. Part of the House of Sak, built on a $0 budget. Read the story →

---

What it is

2,003 training + 113 held-out test examples in OpenAI chat JSONL format, teaching small models (0.5B–7B) when to call a tool, how to carry multi-turn context, when to answer directly, and when to refuse.

Each example is a ChatML-formatted conversation with tool schemas in the system prompt:

{"messages": [
  {"role": "system", "content": "<tools>[...]</tools>"},
  {"role": "user", "content": "What's the weather in Tokyo?"},
  {"role": "assistant", "content": "<tool_call>\n{\"name\": \"get_weather\", \"arguments\": {\"location\": \"Tokyo\"}}\n</tool_call>"}
]}

Quick start

from datasets import load_dataset

# Training split
ds = load_dataset("Nanthasit/sakthai-combined-v6", split="train")
print(len(ds))  # 2003

# Held-out test split (113 examples)
ds_test = load_dataset("Nanthasit/sakthai-combined-v6", split="test")
print(len(ds_test))  # 113

Pair with the irrelevance-supplement (10 "don't-call" examples) for stronger irrelevance behaviour.

With tool rendering

from datasets import load_dataset
from transformers import AutoTokenizer
import json

tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
ds = load_dataset("Nanthasit/sakthai-combined-v6", split="train")

def render(rec):
    msgs = rec.get("messages") or []
    fixed = []
    for m in msgs:
        m = dict(m)
        if m.get("tool_calls"):
            calls = []
            for t in m["tool_calls"]:
                fn = dict(t["function"])
                if isinstance(fn.get("arguments"), str):
                    fn["arguments"] = json.loads(fn["arguments"])
                calls.append({**t, "function": fn})
            m["tool_calls"] = calls
        fixed.append(m)
    return tok.apply_chat_template(fixed, tools=rec.get("tools") or None, tokenize=False)

texts = [render(r) for r in ds.select(range(3))]
for t in texts:
    print(t[:200] + "...")

---

Data Quality Card

Dataset Composition

Metric	Value
Total rows (train)	2,003
Total rows (test)	113
Total messages	8,922 (avg 4.5/row)
Tool call instances	2,189
Rows with tool schemas	1,913 (96%)
Multi-turn examples (2+ user turns)	179
Format integrity	✅ All rows have messages, tools, conversations keys
Null messages	0

Role Distribution

Role	Count
system	1,803
user	2,079
assistant	3,549
tool	1,491

Category Breakdown

Category	~Count	Purpose
Tool-calling (single)	1,380	<tool_call> with get_weather, search_web, calculate, etc.
Multi-turn	179	Follow-ups with context carry-over over 2+ user turns
Edge cases	200	Ambiguous / partial / hallucination-prone prompts
Energy-aware	50	Charge-state-conditioned responses
Irrelevance	50	General Q&A needing no tool call
Safety / rejection	73	Harmful-prompt refusals + jailbreak attempts
Tool-calling (parallel)	~71	Two or more tools called in a single turn

Evolution from v5

Version	Examples	Change
v1	500	Initial tool-calling seeds
v2	800	Multi-turn dialogues
v3	1,000	Irrelevance detection
v4	1,150	Safety rejections
v5	1,280	Quality filtering + dedup
v6	2,003	Edge cases + expanded tool examples + safety prompts

Tool Schema Coverage

~20 unique tool schemas across these categories:

• Built-in Agent: learn, recall, search, read_file, forget, send_message
• Knowledge/Info: get_weather, get_current_time, calculate, search_web, get_stock_price, translate_text
• Utility: set_reminder, save_note, send_email, get_fibonacci, get_fact, check_palindrome, sum_numbers

---

Trained models

Powers the whole SakThai family — flagship context-1.5b-merged, 0.5b / 7b variants, the tool-calling LoRAs, and coder-1.5b.

13 models · 6 datasets · 3 Spaces — full collection →

Sibling datasets

Dataset	Purpose	Downloads
sakthai-combined-v7	🆕 Latest tool-calling data (2,309 ex., 86 tools)	0
sakthai-irrelevance-supplement	Safety supplement — when NOT to call tools	0
sakthai-kaggle-notebooks	Training notebooks & deploy scripts	103
SimpleToolCalling	Early experiment (deprecated)	52

Limitations

• Synthetic: Generated data, not collected from real user interactions
• Domain-specific: Tools reflect SakThai-Agent capabilities
• Language: English only
• Size: 2,003 training examples — moderate; supplement with v7 for production use
• v6 → v7: This dataset is superseded by combined-v7 (2,309 rows, 86 tools, better safety coverage, 100% messages format). New projects should prefer v7.

License

Apache 2.0 — the dataset was originally published with other license but is made available under Apache 2.0 for maximum compatibility.