---
license: gemma
base_model: unsloth/functiongemma-270m-it
library_name: peft
pipeline_tag: text-generation
tags:
- function-calling
- tool-use
- ethereum
- wallet
- lora
- sft
- gguf
datasets:
- ef-dai-team/wallet-tool-calling-ft
---

# functiongemma-270m-wallet-ft

A LoRA fine-tune of `unsloth/functiongemma-270m-it` intended to turn a
natural-language wallet request into the exact structured tool call a macOS
Ethereum wallet can execute.

## ⚠️ This fine-tune did not work

On the 307-case evaluation set it scores **8.8%**, against **8.1%** for the
untuned base model. Every one of its 27 passing cases is a case where the correct
answer is *not* to emit a tool call:

| Category | Cases | Score |
| --- | --- | --- |
| transfers | 101 | **0%** |
| swaps | 96 | **0%** |
| multi-turn | 75 | **0%** |
| ablation (clarifying question expected) | 28 | 82.1% |
| refusal (no call expected) | 7 | 57.1% |

Asked to "Send 0.1 ETH to vitalik.eth" it emits `executeTx` with hallucinated
argument names (`afterQuotes`, `callerName`, `callerString`). Fine-tuning moved
the core task by nothing.

**Do not deploy this.** It is published as a reproducible negative result.

## What worked instead

The same dataset and the same LoRA recipe take Gemma-4 E4B from 9.8% to
**80.1%** — see [`ef-dai-team/gemma-4-E4B-wallet-ft`](https://huggingface.co/ef-dai-team/gemma-4-E4B-wallet-ft).
The binding constraint is model capacity, not training data.

## Contents

- `functiongemma-270m-wallet-ft.Q8_0.gguf` — merged + quantised, 291 MB, runs via
  `llama-cpp-python`
- `adapter/` — the LoRA adapter (r=16, α=16) and training state

## Training data

[`ef-dai-team/wallet-tool-calling-ft`](https://huggingface.co/datasets/ef-dai-team/wallet-tool-calling-ft)
— 1739 SFT examples, disjoint from the evaluation set by construction.

## Evaluation

Scored by a deterministic binary scorer: a case passes only if every field of
every emitted call matches gold exactly. Full per-case results, including this
model's actual output on all 307 cases, are in the
[eval report Space](https://huggingface.co/spaces/ef-dai-team/wallet-tool-calling-eval).

## License

`gemma` — inherited from the base model [`unsloth/functiongemma-270m-it`](https://huggingface.co/unsloth/functiongemma-270m-it). Use is governed by the [Gemma Terms of Use](https://ai.google.dev/gemma/terms).
The training data is Apache-2.0 and licensed separately.