ClarkBear/gemma4-e2b-mobile-actions-200-lora

Gemma 4 E2B Mobile Actions LoRA, 200-Sample Fine-Tune

Main model card: google/gemma-4-E2B-it

This repository provides the LoRA adapter for an experimental Gemma 4 E2B fine-tune on google/mobile-actions.

The adapter was trained locally on an Apple M1 Pro and can be loaded on top of google/gemma-4-E2B-it.

Performance

The merged model produced from this adapter was evaluated on 200 held-out examples from google/mobile-actions (metadata == "eval").

Metric	Score
Format valid rate	94.0%
Function name accuracy	94.0%
Required arguments present	92.0%
Exact match	85.0%

Per-Function Results

Function	N	Format	Name	Required	Exact
create_contact	43	88.4%	88.4%	88.4%	86.0%
create_calendar_event	42	90.5%	90.5%	81.0%	78.6%
show_map	38	97.4%	97.4%	97.4%	68.4%
open_wifi_settings	20	90.0%	90.0%	90.0%	90.0%
send_email	20	100.0%	100.0%	100.0%	95.0%
turn_on_flashlight	20	100.0%	100.0%	100.0%	100.0%
turn_off_flashlight	17	100.0%	100.0%	100.0%	100.0%

Training

Setting	Value
Base model	google/gemma-4-E2B-it
Dataset	google/mobile-actions
Training samples	200
Epochs	1
LoRA rank / alpha	16 / 32
Batch / accumulation	1 / 16
Max sequence length	1024
Precision	bfloat16
Hardware	Apple M1 Pro, MPS
Training time	35m47s including merge

Usage

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoProcessor

base_model = "google/gemma-4-E2B-it"
adapter_id = "YOUR_USERNAME/gemma4-e2b-mobile-actions-200-lora"

processor = AutoProcessor.from_pretrained(adapter_id)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter_id)

Examples

This repository includes a runnable PEFT example:

python examples/run_lora.py \
  --adapter-id ClarkBear/gemma4-e2b-mobile-actions-200-lora \
  --prompt "Turn on the flashlight"

You need Hugging Face access to google/gemma-4-E2B-it.

Limitations

This is an experimental small-data LoRA adapter. It was trained on only 200 examples and should be evaluated before use in production or on-device workflows.