Nanthasit/sakthai-toolcalling-v1

SakThai Tool-Calling v1

Supervised fine-tuning data for teaching a small LLM (Qwen2.5-0.5B family) to call SakThai-Agent's tools, plus general function-calling competence.

Each row is a single-turn example in OpenAI/Qwen messages + tools format. Render for training with tokenizer.apply_chat_template(messages, tools=tools, ...).

Composition

Source	Rows	Notes
Synthetic (SakThai)	~83	All 8 BUILTIN_TOOLS (learn, recall, search, forget, read_file, run_command, send_telegram_message, run_agent_loop) + 6 no-tool negatives.
Public (glaive)	~159	First tool-call turn from glaiveai/glaive-function-calling-v2, reformatted.

Deduplicated by (user message, tool name) -> 167 unique (151 train / 16 test).

Quick Start (TRL SFT)

from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer
from trl import SFTConfig, SFTTrainer

ds = load_dataset("Nanthasit/sakthai-toolcalling-v1", split="train")
base = "Qwen/Qwen2.5-0.5B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base)

trainer = SFTTrainer(
    model=model, processing_class=tok, train_dataset=ds,
    args=SFTConfig(output_dir="sakthai-toolcall-sft", num_train_epochs=3,
                   per_device_train_batch_size=4, learning_rate=2e-4,
                   bf16=True, max_length=2048),
)
trainer.train()

Quick Start (inference)

from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Nanthasit/sakthai-persona-0.5b-lora"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

row = load_dataset("Nanthasit/sakthai-toolcalling-v1", split="test")[0]
inputs = tok.apply_chat_template(row["messages"], tools=row["tools"],
                                  add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(inputs, max_new_tokens=256)[0][inputs.shape[-1]:], skip_special_tokens=True))

Limitations

• Small (v1). Single-turn only - no tool-result -> follow-up turns yet.
• Synthetic SakThai args are templated.
• Public rows use generic tools (weather, stock, etc.).

Roadmap (v2)

• Multi-turn tool-result -> follow-up assistant turns.
• Negative sampling for not-calling-tool.
• More SakThai-specific tooling (Composio / Hermes / smolagents).
• Larger public mixture (Salesforce/xlam-function-calling-60k slice).

Built for fine-tuning Nanthasit/sakthai-persona-0.5b-lora / Qwen2.5-0.5B-Instruct.

Citation

@misc{sakthai-toolcalling-v1,
  title  = {SakThai Tool-Calling v1},
  author = {Nanthasit},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Nanthasit/sakthai-toolcalling-v1}
}