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

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

Model description

GGUF conversion of Qrzysztof/functiongemma-270m-it-prepaid-cards-v2 (SafeTensors), a google/functiongemma-270m-it fine-tune that buys prepaid cards, checks balances, and lists transactions via tool calls in 107 languages with realistic noisy/multi-turn input.

Runnable with llama.cpp and any GGUF-compatible runtime (llama-cpp-python, LM Studio, Ollama, …).

Files

File	Size	Description
model-f16.gguf	551 MB	Full-precision weights
model-Q8_0.gguf	300 MB	8-bit quantization — low-end hardware (see below)

How to use

# CLI (llama.cpp)
llama-cli -m model-Q8_0.gguf -cnv -p "I want to buy a $20 card"

# llama-cpp-python — feed the FunctionGemma-rendered prompt
from llama_cpp import Llama

llm = Llama(model_path="model-Q8_0.gguf", n_ctx=4096, n_gpu_layers=0, verbose=False)
prompt = "<bos><start_of_turn>developer\nYou are a model that can do function calling with the following functions<start_function_declaration>..."  # tokenizer.apply_chat_template(messages, tools=tools, add_generation_prompt=True)
print(llm(prompt, max_tokens=96, temperature=0.0)["choices"][0]["text"])
# <start_function_call>call:purchase_card{"amount": 20, "card_type": "digital_prepaid_visa", ...}<end_function_call>...

Note: llama-cli (master) had a REPL quirk when passing prompts with -p/-f at the time of writing; llama-cpp-python (same backend) is the tested path.

Intended uses & limitations

Same as the parent model (see the SafeTensors card): synthetic data, uneven language quality, no backend — plus the usual quantization caveats below.

How it was made

# 1. conversion (llama.cpp convert_hf_to_gguf.py; needs the llama.cpp repo layout)
python3 convert_hf_to_gguf.py <hf_model_dir> --outfile model-f16.gguf --outtype f16
# 2. 8-bit quantization
llama-quantize model-f16.gguf model-Q8_0.gguf Q8_0

Gemma3 270M gotcha: the HF tokenizer reports 262,146 tokens (two model-specific vision tokens added at load). convert_hf_to_gguf.py asserts max(vocab) < vocab_size and fails. Fix: in a conversion copy, drop the *_token / model_specific_special_tokens keys from tokenizer_config.json and prune added tokens with id ≥ 262144 from tokenizer.json → exactly 262,144 tokens, identical text behavior.

Evaluation

Same prompts & greedy decoding as the other formats, over the held-out v2 test subset (N=40).

Format	Success rate
SafeTensors (reference)	40/40 = 100%
GGUF f16	TODO
GGUF Q8_0	40/40 = 100%

Q8_0 shows no measurable drop on tool-name selection in this subset; arg wording can differ slightly from f16 in rare cases.

8-bit / low-end hardware notes

• Q8_0 file is 300 MB; runtime RSS ≈ 350–450 MB with a 4K context → runs on 2 GB-RAM devices; CPU-only inference works (slow but usable) on Raspberry-Pi-class hardware.
• Want even smaller? llama-quantize model-f16.gguf model-Q4_K_M.gguf Q4_K_M (≈170 MB) trades a bit more accuracy.

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: MLX 8-bit · ONNX
• Dataset: ...-tool-calling-v2