Supervised fine-tune of meta-llama/Llama-3.2-1B-Instruct for function calling, trained on the Salesforce xLAM dataset. This is the v1 SFT model in a four-model study comparing SFT, data scaling, GRPO, and supervised refusal training for small-model function calling.

Results (BFCL v4)

Evaluated on the Berkeley Function Calling Leaderboard v4, compared to the base model:

Category	Base	This model (v1 SFT)
simple_python	75.0	77.5
multiple	50.5	74.0
live_simple	31.8	57.0
live_multiple	7.3	38.8
live_relevance	43.8	93.8
live_irrelevance	67.3	16.9
irrelevance	35.8	5.8
parallel	44.0	1.0
parallel_multiple	15.0	2.0

Large gains on call-required categories (+23–50 pts). Note the irrelevance regression (35.8→5.8): training on single-call xLAM data — which contains no "no-call" examples — teaches the model to always call a function. This regression is addressed in the companion Path B model.

Training details

Base: meta-llama/Llama-3.2-1B-Instruct
Data: Salesforce/xlam-function-calling-60k, filtered to 27k single-call examples
Method: LoRA (r=16, α=32, dropout=0.05) on all attention + MLP projections
Config: 1 epoch, effective batch 16, lr 2e-4 cosine, bf16
Hardware: 1× A100, ~55 min
Final eval loss: 0.11 · token accuracy: 97.4%

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained("Keitsuna123/llama-3.2-1b-fc-sft-full")
model = AutoModelForCausalLM.from_pretrained(
    "Keitsuna123/llama-3.2-1b-fc-sft-full", torch_dtype=torch.bfloat16, device_map="auto"
)

messages = [{"role": "user", "content": "What's the weather in Tokyo?"}]
tools = [{"type": "function", "function": {
    "name": "get_weather", "description": "Get the weather for a location",
    "parameters": {"type": "object", "properties": {"location": {"type": "string"}}, "required": ["location"]}
}}]

text = tokenizer.apply_chat_template(messages, tools=tools, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=128, do_sample=False)
print(tokenizer.decode(out[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True))
# {"name": "get_weather", "parameters": {"location": "Tokyo"}}

Limitations

Irrelevance detection regressed vs. the base model (see results). Use the Path B variant if refusal behavior matters.
Single-call only. Parallel function calls are near-zero — Llama 3.2's chat template supports one tool call per turn, and training was filtered accordingly.

Part of a series

Model	Description
fc-sft-full	v1 SFT on xLAM (this model)
fc-sft-v2-merged	SFT on xLAM + distilabel (data-scaling ablation)
fc-grpo	GRPO for irrelevance recovery (negative result)
fc-pathb	Supervised refusal training (effective irrelevance fix)

References & Acknowledgements

This work builds on:

Base model: Llama 3.2 (Meta AI) — meta-llama/Llama-3.2-1B-Instruct
Training data: Salesforce xLAM — xlam-function-calling-60k
Benchmark: Berkeley Function Calling Leaderboard (BFCL) — Gorilla project
Frameworks: HuggingFace TRL, PEFT, Transformers

Citation

bibtex

@misc{taketsuna2026_fc_smallmodel,
  title  = {Small-Model Function Calling: Comparing SFT, Data Scaling, GRPO, and Supervised Refusal Training},
  author = {Taketsuna, Keiichi},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/Keitsuna123}},
  note   = {Llama-3.2-1B function-calling post-training study on BFCL v4}
}