llama-3.2-1b-fc-dpo/README.md

base_model, library_name, tags, license, language

base_model	library_name	tags	license	language
meta-llama/Llama-3.2-1B-Instruct	transformers	function-calling tool-use llama dpo preference-optimization bfcl	llama3.2	en

Llama-3.2-1B-Instruct — Function Calling (DPO)

The v1 SFT model further trained with DPO (Direct Preference Optimization) to recover an irrelevance-detection regression. DPO is a preference optimization method — not reinforcement learning — that optimizes a model to prefer "chosen" over "rejected" responses directly, without a reward model or sampling. This is one of six models in a study comparing supervised, preference, and RL post-training for small-model function calling.

Code & writeup: https://github.com/keitake123/llama-function-calling-study

Results (BFCL v4)

Category	Base	v1 SFT	GRPO (RL)	DPO (pref.)	Path B (SFT)
irrelevance	35.8	5.8	5.4	8.3	21.7
live_irrelevance	67.3	16.9	17.2	28.6	36.8
simple_python	75.0	77.5	77.2	77.5	78.2
multiple	50.5	74.0	74.0	73.5	72.0
live_simple	31.8	57.0	56.2	58.5	54.3
live_multiple	7.3	38.8	39.4	37.0	35.7
live_relevance	43.8	93.8	93.8	81.2	81.2

Finding: DPO produced partial irrelevance recovery (5.8→8.3 non-live; 16.9→28.6 live) — clearly better than GRPO (which showed no recovery) but below refusal-SFT. This fits a clean pattern across the three methods:

Recovery scaled with how directly each method exposed the model to the target behavior. GRPO never samples a refusal (the SFT always-call prior is too strong) → no signal → no recovery. DPO is shown refusals as the "chosen" response → partial recovery, at low cost to call categories. Refusal-SFT trains directly on refusals → strongest recovery.

Ordering: imitation (SFT) > preference (DPO) > RL (GRPO). RL recovers least because it cannot reinforce a behavior absent from the policy's sampling distribution.

Notably, DPO preserved call-required categories well (simple_python, live_simple held up), giving arguably the best precision/recall balance despite a lower raw irrelevance score than Path B.

Training details

• Base: v1 SFT model, cleanly merged (see note below)
• Method: DPO (TRL DPOTrainer), LoRA (r=16, α=32)
• Data: ~1,600 preference pairs, 50/50 mix:
  • Irrelevance pairs: chosen = natural refusal, rejected = v1's hallucinated call (generated by running v1 on the irrelevance prompts)
  • Call-required pairs: chosen = correct call, rejected = wrong-function call
• Config: 1 epoch, effective batch 8, lr 5e-6, beta (KL) 0.3, bf16
• Hardware: 1× RTX A6000

Note on stability: an initial run with beta=0.1 and a stacked (unmerged) reference adapter caused output-format drift (the model emitted a malformed call schema and occasionally code). Fixed by (1) cleanly merging the v1 adapter into the base before DPO so the reference model is well-formed, and (2) raising beta to 0.3 for a stronger anchor. See the repo's LESSONS.md.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained("Keitsuna123/llama-3.2-1b-fc-dpo")
model = AutoModelForCausalLM.from_pretrained(
    "Keitsuna123/llama-3.2-1b-fc-dpo", 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))

Limitations

• Residual irrelevance failures concentrate in short, factual-seeming queries ("what's 2+2", "capital of France") — the same sub-case that beat every method in this study.
• Parallel calls near-zero (single-call training only).

Part of a series

Model	Method	Description
fc-sft-full	SFT	v1 SFT on xLAM
fc-sft-v2-merged	SFT	+ distilabel (data-scaling ablation)
fc-grpo	RL	GRPO (no recovery)
fc-pathb	SFT	refusal-SFT (best recovery)
fc-dpo	DPO	preference optimization (this model)

References & Acknowledgements

• Base model: Llama 3.2 (Meta AI) — meta-llama/Llama-3.2-1B-Instruct
• Training data: Salesforce xLAM — xlam-function-calling-60k; preference pairs constructed for this study
• Method: DPO (Rafailov et al., "Direct Preference Optimization")
• Benchmark: Berkeley Function Calling Leaderboard (BFCL) — Gorilla project
• Frameworks: HuggingFace TRL (DPOTrainer), PEFT, Transformers

Citation

@misc{taketsuna2026_fc_smallmodel,
  title  = {Small-Model Function Calling: Comparing SFT, Data Scaling, GRPO, and DPO Post-Training},
  author = {Taketsuna, Keiichi},
  year   = {2026},
  howpublished = {\url{https://github.com/keitake123/llama-function-calling-study}},
  note   = {Llama-3.2-1B function-calling post-training study on BFCL v4}
}