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

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

Model description

ONNX Runtime export of Qrzysztof/functiongemma-270m-it-prepaid-cards-v2 (a google/functiongemma-270m-it fine-tune for prepaid-card tool calling in 107 languages with noisy/multi-turn input) — for CPU, mobile, and web-browser inference (ONNX Runtime Web / Transformers.js).

Files

File	Size	Description
model.onnx + model.onnx.data	1.07 GB	fp32 export (exact reference)
model-fp16.onnx + model-fp16.onnx.data	536 MB	fp16 export — recommended for browsers

Graph

Full-context decoder (no KV-cache inputs): every step feeds the whole context and returns full logits.

Input	Shape	Dtype
input_ids	[batch, seq] dynamic	int64
attention_mask	[batch, seq] dynamic	int64
position_ids	[batch, seq] dynamic	int64
logits (output)	[batch, seq, vocab] dynamic	fp32 / fp16

How to use

Python (onnxruntime)

import numpy as np, onnxruntime as ort, torch
from transformers import AutoTokenizer

sess = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])
tokenizer = AutoTokenizer.from_pretrained("Qrzysztof/functiongemma-270m-it-prepaid-cards-v2")

prompt = tokenizer.apply_chat_template(messages, tools=tools, add_generation_prompt=True, tokenize=False)
ids = tokenizer(prompt, return_tensors="pt").input_ids
for _ in range(48):
    logits = sess.run(None, {
        "input_ids": ids.numpy().astype(np.int64),
        "attention_mask": np.ones((1, ids.shape[1]), dtype=np.int64),
        "position_ids": np.arange(ids.shape[1], dtype=np.int64).reshape(1, -1),
    })[0]
    ids = torch.cat([ids, torch.tensor([[logits[0, -1].argmax()]])], dim=1)
print(tokenizer.decode(ids[0], skip_special_tokens=False))

Web browser (ONNX Runtime Web / Transformers.js)

import { AutoTokenizer, AutoModelForCausalLM } from "@huggingface/transformers";
const tokenizer = await AutoTokenizer.from_pretrained("Qrzysztof/functiongemma-270m-it-prepaid-cards-v2-onnx");
const model = await AutoModelForCausalLM.from_pretrained("Qrzysztof/functiongemma-270m-it-prepaid-cards-v2-onnx",
                                                         { dtype: "fp16", device: "wasm" });

The graph uses only standard ONNX ops (Gemm/MatMul/Add/Softmax/…), so it runs under the WASM backend. Full-context decode is O(seq²) — fine for a 270M model with short tool-call outputs.

Intended uses & limitations

Same as the parent model (see the SafeTensors card): synthetic data, uneven language quality, no backend. Performance note: fp32 CPU decode of this graph is slow (no KV cache); prefer fp16 in browsers and GGUF/MLX for interactive local use.

How it was made

optimum-onnx is pinned to transformers <4.58 (incompatible with transformers 5.x), so the export uses torch.onnx.export directly:

python3 export_onnx.py --model-dir <hf_model_dir> --out model.onnx             # fp32
python3 export_onnx.py --model-dir <hf_model_dir> --out model-fp16.onnx --dtype fp16

Settings: eager attention, use_cache=False, opset 17, dynamic axes.

Correctness: verified token-identical greedy decoding vs the PyTorch reference (export_onnx.py --check).

Evaluation

Same prompts & greedy decoding as the other formats, over a held-out v2 test subset (N=20, 32 max tokens — full-context CPU decode is slow).

Format	Success rate
SafeTensors (reference)	40/40 = 100% (N=40)
ONNX fp32 (ORT CPU)	20/20 = 100%
ONNX fp16 (graph)	same graph semantics; tested identical logits in spot checks

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

• Parent: ...-v2 (SafeTensors)
• Siblings: GGUF f16+Q8_0 · MLX 8-bit
• Dataset: ...-tool-calling-v2