strdst77/TinyToolCaller

🛠 Hugging Face Model Card Template

Model Overview

TinyToolCaller is a fine‑tuned tool‑calling model trained on synthetic datasets for structured tool use. It uses LoRA adapters for efficient training.

Training Environment

GPU	VRAM	PyTorch	TRL	PEFT	bitsandbytes	Wall‑clock (W&B)	Peak memory (W&B)	McNemar p‑value
Tesla T4	15 GB	2.11	1.10	0.20	0.50	X hrs	Y GB	p = 0.0ZZ

(Replace X/Y/p with your actual W&B values and test result.)

Datasets

• Synthetic tool‑calling dataset (hf-dataset folder)
• Optional GSM8K evaluation (1319 examples)

Training Procedure

• Supervised fine‑tuning (SFT) with max sequence length 512
• Optimizer: AdamW
• Learning rate: 2e‑5
• LoRA rank: 16, alpha: 32, dropout: 0.05

Evaluation

• GSM8K accuracy: XX%
• McNemar test: p = 0.0ZZ (paired significance)

Limitations

• Optimized for tool‑calling tasks, not general reasoning.
• Performance outside structured tool use may degrade.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("strdst7/TinyToolCaller-T4")
tokenizer = AutoTokenizer.from_pretrained("strdst7/TinyToolCaller-T4")

inputs = tokenizer("Call weather API for Kuala Lumpur", return_tensors="pt")
outputs = model.generate(**inputs)
print(tokenizer.decode(outputs[0]))

Citation

If you use this model, please cite:

@misc{tinytoolcaller2026,
  author = {Astrid},
  title = {TinyToolCaller},
  year = {2026},
  publisher = {Hugging Face},
}

---