LFM2.5-1.2B-Instruct-mobile-actions

LFM2.5-1.2B-Instruct-mobile-actions

A LoRA fine-tune of LiquidAI/LFM2.5-1.2B-Instruct on google/mobile-actions for on-device function calling on Android.

The model translates natural language instructions into executable function calls for Android OS system tools (e.g. sending messages, setting alarms, making calls, web search).

Training Details

LoRA Configuration

Training Results

Training was response-only (user/system tokens masked from loss). Eval loss tracks train loss closely with no overfitting.

Inference Results

Evaluated on 20 held-out examples from the google/mobile-actions eval split: 20/20 correct (100% accuracy)

*First call includes CUDA warmup. Typical inference: ~1-5s on NVIDIA L4.

Accuracy by function type:

Usage

With PEFT + Transformers

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained("LiquidAI/LFM2.5-1.2B-Instruct")
model = PeftModel.from_pretrained(base_model, "kshitijthakkar/LFM2.5-1.2B-Instruct-mobile-actions")
tokenizer = AutoTokenizer.from_pretrained("kshitijthakkar/LFM2.5-1.2B-Instruct-mobile-actions")

messages = [
    {"role": "system", "content": "You are a helpful assistant with access to the following functions. Use them if required.\n\n" + tools_json},
    {"role": "user", "content": "Send a text message to John saying I'll be late"}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs, max_new_tokens=256, temperature=0.1, top_k=50, top_p=0.1, repetition_penalty=1.05)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With Unsloth (faster inference)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="kshitijthakkar/LFM2.5-1.2B-Instruct-mobile-actions",
    max_seq_length=2048,
)
FastLanguageModel.for_inference(model)

Dataset Format

Each training example contains:

• System prompt: Available Android function definitions (JSON)
• User message: Natural language instruction (e.g. "Set an alarm for 7am tomorrow")
• Assistant response: Function call as JSON (e.g. [{"name": "set_alarm", "arguments": {"datetime": "..."}}])

The model was trained on responses only, learning to map user instructions to the correct function calls given available tool definitions.

Intended Use

• On-device function calling for Android system tools
• Translating natural language to structured API calls
• Edge deployment for mobile assistants

Limitations

• Trained only on Android system tool functions from the google/mobile-actions dataset
• May not generalize to arbitrary function schemas outside the training distribution
• Inherits limitations of the base LFM2.5-1.2B-Instruct model

Training Script

The training script is available at: github.com/Mandark-droid/LFM2.5-1.2B-Instruct-mobile-actions