shawneil/NanoMind-MobileActions

NanoMind — Mobile Actions (Function Calling)

Fine-tuned on google/mobile-actions for on-device function calling (calendar, email, contacts, maps, flashlight, Wi-Fi).

Quick Start

from hf_wrapper import NanoMindForFunctionCalling
import tiktoken, torch, json

model = NanoMindForFunctionCalling.from_pretrained(
    "shawneil/NanoMind-MobileActions"
)
model.eval()
enc = tiktoken.get_encoding("gpt2")

# Build a prompt in the same format used during training
tools_json = json.dumps([{"function": {"name": "create_calendar_event",
    "description": "Creates a calendar event.",
    "parameters": {"properties": {"title": {"type": "STRING"},
    "datetime": {"type": "STRING"}}, "required": ["title","datetime"]}}}])

prompt = (
    "<|tools|>\n" + tools_json + "\n<|endtools|>\n"
    "<|system|>\nYou are a function-calling assistant.\n<|endoftext|>\n"
    "<|user|>\nSet a meeting tomorrow at 3 PM called 'Budget Review'.\n<|endoftext|>\n"
    "<|assistant|>\n"
)
ids = torch.tensor([enc.encode_ordinary(prompt)], dtype=torch.long)
out = model.generate_text(ids, max_new_tokens=80, temperature=0.0, top_k=1)
print(enc.decode(out[0].tolist()))

Training details

• Base: NanoMind SFT-Dolly (shawneil/NanoMind-SFT-Dolly)
• Dataset: google/mobile-actions (~9.65k rows)
• Loss only on assistant/tool-call tokens
• fp16 · torch.compile · 2× T4 DDP