Nanthasit/sakthai-combined-v7

SakThai Combined Dataset v7

Tool-calling, multi-turn, safety, and edge cases — the base training data for the SakThai model family.

Part of the SakThai Model Family Collection.

Dataset Statistics

Property	Value
Train examples	2,309
Test examples	115
Total	2,424
Format	OpenAI chat JSONL
Languages	English, Thai
License	Apache 2.0
Tools	86 unique function schemas
Files	2 JSONL files (train + test)

Category Breakdown (train)

Category	Count	Description
Tool-calling	~900	Single and multi-tool invocations
Multi-turn	~500	3+ message exchanges
Irrelevance	~300	Requests needing no tools
Safety	~200	Refusal, guardrails, out-of-scope
Bilingual (Thai)	~100	Thai language tool-calling

Splits

Split counts verified via Datasets Server:

Split	Config	Rows	Parquet Size	Columns	Description
train	train	2,309	3.9 MB	2 (messages, tools)	Tool-calling training examples
train	test	115	340 KB	2 (messages, tools)	Held-out test set
Total		2,424	4.3 MB

Note: Both configs have split name train (the Hugging Face Datasets Server convention).
The dataset viewer is fully enabled — use /first-rows, /statistics, /search, and /filter endpoints to explore without downloading.

Message Length Distribution (train)

• Mean messages per example: 5.1 ± 2.8 (min: 3, max: 23)
• Mean tools per example: 4.3 ± 4.6 (min: 0, max: 24)
• 19 rows (0.8%) have null tools (irrelevance examples with no schema)

Data from /statistics endpoint.

Data Fields

Each JSONL row has a messages field; a tools field is present when function definitions are offered:

Field	Type	Description
messages	list[dict]	Chat conversation turns following the OpenAI messages format (always present)
tools	list[dict]	Available function definitions the assistant may call (optional — absent on no-tool/irrelevance rows; empty list on a subset of rows)

Messages Structure

Each entry in messages has these fields:

Field	Type	Required	Description
role	string	✅	One of system, user, assistant, or tool
content	string	varies	Message body text. May be null when tool_calls is present
tool_calls	list[dict]	❌	Present only on assistant messages that invoke a function
tool_call_id	string	❌	Present only on tool messages; links back to the call

Tool Calls Structure

Each entry in tool_calls:

Field	Type	Description
id	string	Unique call ID (e.g. call_abc123)
type	string	Always "function"
function.name	string	Name of the function to invoke
function.arguments	string	JSON-stringified argument object

Tools Structure

Each entry in tools:

Field	Type	Description
type	string	Always "function"
function.name	string	Unique function name
function.description	string	Natural language description of what the function does
function.parameters	dict	JSON Schema object defining accepted parameters and their types

Data Format

{
  "messages": [
    {"role": "system", "content": "You are SakThai-Agent..."},
    {"role": "user", "content": "Search for papers on RLHF"},
    {"role": "assistant", "tool_calls": [
      {"function": {"name": "search_papers", "arguments": "{\"query\": \"RLHF\"}"}}
    ]},
    {"role": "tool", "content": "{\"results\": [...]}"},
    {"role": "assistant", "content": "Here's what I found..."}
  ],
  "tools": [
    {"type": "function", "function": {
      "name": "search_papers",
      "description": "Search academic papers",
      "parameters": {"type": "object", "properties": {...}}
    }}
  ]
}

Sample Row

{
  "messages": [
    {"role": "user", "content": "What's the weather in Bangkok?"}
  ],
  "tools": [{"type": "function", "function": {
    "name": "get_weather",
    "description": "Get current weather",
    "parameters": {
      "type": "object",
      "properties": {
        "location": {"type": "string"},
        "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}
      },
      "required": ["location"]
    }
  }}]
}

Usage

from datasets import load_dataset

# Load training split (config name required — repo has train/test configs)
train = load_dataset("Nanthasit/sakthai-combined-v7", name="train", split="train")
print(f"Train: {len(train)} examples")

# Load test split
test = load_dataset("Nanthasit/sakthai-combined-v7", name="test", split="train")
print(f"Test: {len(test)} examples")

# Format for TRL training
def to_text(ex):
    from transformers import AutoTokenizer
    tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
    return {"text": tokenizer.apply_chat_template(
        ex["messages"], tools=ex.get("tools"), tokenize=False
    )}

Loading with Pandas

import pandas as pd

train_df = pd.read_json("hf://datasets/Nanthasit/sakthai-combined-v7/data/train.jsonl", lines=True)
print(train_df.shape)  # (2309, 2)

Data Quality

• Validation: Each example checked against tool schemas for correctness
• Coverage: 86 unique tool schemas across diverse domains
• Edge cases: Empty strings, null values, unicode, extreme lengths
• Irrelevance: ~300 examples where model must NOT call tools

Recommended: Use v10 (normalized v7 + v8)

For optimal coverage with consistent parameter names, use sakthai-combined-v10 — a normalized version of v7 + v8:

from datasets import load_dataset

combined = load_dataset("Nanthasit/sakthai-combined-v10", split="train")
print(f"Combined: {len(combined)} examples")

This gives you the broadest tool schema coverage alongside thoroughly tested edge cases.

Citation

@misc{sakthai-v7-2026,
  author = {Beer (beer-sakthai)},
  title = {SakThai Combined Dataset v7 — Tool-Calling Training Data},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face Datasets},
  howpublished = {\url{https://huggingface.co/datasets/Nanthasit/sakthai-combined-v7}}
}

Related

• v6 (predecessor)
• Benchmark v2
• Irrelevance Supplement
• SakThai Model Family