Qrzysztof/functiongemma-270m-it-prepaid-cards-v2

FunctionGemma 270M IT — Prepaid Cards Tool-Calling (v2, SafeTensors)

Model description

A fine-tuned version of google/functiongemma-270m-it (Gemma 3 270M, 268M params) that recognizes prepaid-card intents in chat and emits the correct tool call:

Tool	Purpose
purchase_card(amount, card_type, email?, currency?)	Buy a Digital Prepaid Visa or Virtual Prepaid Mastercard
get_card_balance(card_number)	Check the balance of a card
get_transaction_history(card_number, limit?)	List a card's transactions

Trained on the v2 dataset: 107 languages, multi-turn conversations (card number in one message, request in another; clarification loops; full call→response loops), and realistic user noise (typos, text-speak, dropped articles, scrambled word order) so the model works with how people actually type.

Intended uses & limitations

Intended uses

• Chat agents that buy prepaid cards, answer balance questions, and show transaction history, in many languages and with noisy/multi-turn input.
• Distillation target: a small model that a backend can drive via the standard FunctionGemma <start_function_call>… protocol.

Limitations & biases

• Synthetic training data. All conversations are generated from hand-written templates; the model has not seen real user traffic.
• Uneven language quality. English and ~30 major languages are the most richly covered; the 20+ low-resource languages were translated by hand and contain approximations. Held-out-language accuracy (89.8% in v1) lags English slightly.
• No backend. The model only emits tool calls; it cannot check balances or buy cards itself.
• Security note: like all small models it can mis-parse card numbers under heavy noise — validate tool arguments before executing payments.
• Gemma license applies (base model license).

How to use

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch, json
from transformers.utils import get_json_schema

model = AutoModelForCausalLM.from_pretrained(
    "Qrzysztof/functiongemma-270m-it-prepaid-cards-v2",
    dtype=torch.bfloat16, attn_implementation="eager")
tokenizer = AutoTokenizer.from_pretrained("Qrzysztof/functiongemma-270m-it-prepaid-cards-v2")

def purchase_card(amount: float, card_type: str, email: str = "", currency: str = "USD") -> str: ...
def get_card_balance(card_number: str) -> str: ...
TOOLS = [get_json_schema(purchase_card), get_json_schema(get_card_balance)]

messages = [
    {"role": "developer", "content": "You are a model that can do function calling with the following functions"},
    {"role": "user", "content": "i wanna buy a 20 dollar card plz"},  # noisy input works
]
inputs = tokenizer.apply_chat_template(messages, tools=TOOLS, add_generation_prompt=True,
                                       return_dict=True, return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(out[0][len(inputs["input_ids"][0]):], skip_special_tokens=False))
# <start_function_call>call:purchase_card{"amount": 20, "card_type": "digital_prepaid_visa", ...}<end_function_call>

Training details

Parameter	Value
Base model	google/functiongemma-270m-it (Gemma 3 270M)
Method	Full fine-tune (all 268M params), TRL SFTTrainer
Data	Qrzysztof/functiongemma-prepaid-cards-tool-calling-v2 — ~1,800 samples/epoch (balanced across 107 languages & intents)
Epochs	3
Batch	8 (T4, bf16, eager attention)
Max length	1024
LR / schedule	5e-5, constant, 50 warmup steps
Hardware	Google Colab T4 GPU

Per-epoch checkpoints: checkpoint/epoch-{1,2,3}.

Evaluation

Method: greedy decoding over the held-out v2 test split (never in training; 5 languages fully held out — ja, ko, ar, sw, ur). A sample counts as correct when the generated text contains the expected tool name and no other tool name (for text-response samples: when it contains no tool call).

Bucket	v1	v2
Overall (295 samples)	91.5% (v1 split)	89.5% (harder v2 split)
purchase_card	92.9%	90.3%
get_card_balance	95.7%	87.0%
get_transaction_history	80.9%	83.8%
Multi-turn chains	98%	100%
Seen languages	94.0%	94.8%
Held-out languages	89.8%	86.0%

Cross-format comparison (subset): torch / GGUF Q8_0 / MLX 8-bit all score 40/40 (100%) on the same 40 prompts; ONNX: see the ONNX repo.

Fine-tuning from this model

This model was fine-tuned with the tutorial below; you can use it as the starting point for a new tool set (or fine-tune google/functiongemma-270m-it directly).

Fine-tuning tutorial

A complete, minimal fine-tune of a FunctionGemma-class model on this data (follows the official FunctionGemma fine-tuning guide).

1. Setup

pip install torch transformers trl datasets accelerate
huggingface-cli login   # accept the gemma license for google/functiongemma-270m-it

2. Load the dataset and normalize messages

The Hub dataset stores messages/tools as JSON strings (Arrow cannot infer the nested schema), and TRL's SFTTrainer needs a uniform struct schema, so normalize first:

import json
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer

def normalize_messages(msgs):
    out = []
    for m in msgs:
        n = {"role": m["role"], "content": m.get("content") or "", "name": None,
             "tool_call_id": m.get("tool_call_id"), "tool_calls": None}
        if m["role"] == "tool":
            n["name"] = m["content"]["name"]
            n["content"] = json.dumps(m["content"]["response"], ensure_ascii=False)
        if m.get("tool_calls"):
            n["tool_calls"] = [{"id": tc.get("id"), "type": tc.get("type", "function"),
                                "function": {"name": tc["function"]["name"],
                                             "arguments": json.dumps(tc["function"]["arguments"], ensure_ascii=False)}}
                               for tc in m["tool_calls"]]
        out.append(n)
    return out

def rows_to_dataset(rows):
    from datasets import Dataset
    return Dataset.from_list([{
        "messages": normalize_messages(r["messages"]),
        "tools": json.dumps(r["tools"], ensure_ascii=False),
    } for r in rows])

ds = load_dataset("Qrzysztof/ecommerce-chat-tool-calling", token=HF_TOKEN)["train"]
train_rows = [{"messages": json.loads(r["messages_json"]), "tools": json.loads(r["tools_json"])}
              for r in ds if r["split"] == "train"]
train_ds = rows_to_dataset(train_rows)

3. Train

import torch
from transformers import AutoModelForCausalLM
from trl import SFTConfig, SFTTrainer

model = AutoModelForCausalLM.from_pretrained("google/functiongemma-270m-it",
                                             dtype=torch.bfloat16, attn_implementation="eager")
tokenizer = AutoTokenizer.from_pretrained("google/functiongemma-270m-it")

trainer = SFTTrainer(
    model=model,
    args=SFTConfig(
        output_dir="functiongemma-ecommerce",
        max_length=1024,          # covers the longest sample + margin
        packing=False,            # keep tool calls intact (no cross-sample packing)
        num_train_epochs=3,
        per_device_train_batch_size=8,
        learning_rate=5e-5,
        lr_scheduler_type="constant",
        warmup_steps=50,
        bf16=True,                # or fp16 on non-Ampere GPUs
        eval_strategy="epoch",
        report_to="none",
    ),
    train_dataset=train_ds,
    processing_class=tokenizer,
)
trainer.train()

TRL applies the FunctionGemma chat template with the per-sample tools column; assistant_only_loss=True (default) masks everything but the model's own turns, so it learns to emit tool calls — not to copy the schema.

4. Evaluate (greedy success rate)

ok = 0
for item in test_rows:
    inputs = tokenizer.apply_chat_template(item["messages"][:-1], tools=item["tools"],
                                           add_generation_prompt=True, return_tensors="pt")
    out = model.generate(**inputs, max_new_tokens=256)
    output = tokenizer.decode(out[0][len(inputs["input_ids"][0]):], skip_special_tokens=False)
    expected = <expected tool name / args from expected_json>
    ok += expected-tool-in-output and no-other-tool-in-output

5. Push

trainer.push_to_hub("YOUR_USER/functiongemma-ecommerce")

Best practices

Data

• Keep noise digit-safe: never corrupt the values the model must extract (prices, ids). The noise.py engine skips any token containing digits.
• Use deterministic train/test splits (by template_id) and hold out whole languages + (for the e-commerce set) whole schemas — that is the only honest way to measure generalization.
• Balance the training subset per (language, intent) — cap the big buckets instead of letting English dominate.

Training

• packing=False for tool-calling data; packed sequences splice mid-call.
• max_length ≥ longest sample + a margin; ~1024 covers these datasets.
• Constant LR + short warmup (the official guide's defaults) work well.
• Upload a checkpoint to the Hub after every epoch — Colab VMs die mid-run, and the last good epoch is always recoverable.

Evaluation

• Always evaluate with greedy decoding for comparability across formats and runs.
• Score two things separately: tool-name selection and argument fidelity (query + every filter key:value pair).
• Compare every exported format (SafeTensors / GGUF / MLX / ONNX) on the same prompts — quantization changes results.

Deployment

• Validate tool arguments server-side before executing anything (a small model can garble a card number under heavy noise).
• In a live agent, follow the FunctionGemma full loop: model call → backend executes → tool response → model continues; never let the model see or emit secrets.
• For browser deployment use the fp16 ONNX file; for low-end hardware the Q8_0 GGUF or MLX 8-bit; for exact reference behavior the SafeTensors model.

Related

• Dataset: v2 · v1
• Formats: GGUF (f16 + Q8_0) · MLX 8-bit · ONNX (fp32/fp16)
• Previous version: v1