Smartwatch LM v0.1
Small GPT-style language model trained for wrist-wearable assistant chat. The model learns conversational replies and emits intent tags such as <INTENT:GET_STEPS> that your app can parse and route to device handlers.
Model details
| Property | Value |
|---|---|
| Architecture | 6-layer causal GPT (~12M params) |
| Context length | 256 tokens |
| Vocab size | 3524 (BPE) |
| Export version | 0.1 |
Files in this repo
| File | Purpose |
|---|---|
smartwatch_lm_merged.onnx | On-device inference (ONNX Runtime, opset 17) |
checkpoint.pt | PyTorch weights for fine-tuning or local chat |
tokenizer.json | BPE tokenizer (Hugging Face Tokenizers format) |
tokenizer_config.json | Tokenizer metadata |
config.json | Model architecture and ONNX I/O names |
model.py | PyTorch model definition + checkpoint loader |
chat.py | Interactive PyTorch chat |
reply_utils.py | Gibberish cleanup, intent parse, slot fill |
onnx_sample.py | ONNX generate + cleanup sample script |
config.py | Generation defaults |
ONNX inference
Inputs: input_ids — int64 tensor, shape [batch, seq] (max seq = 256)
Outputs: logits — float tensor, shape [batch, seq, vocab_size]
Works with ONNX Runtime (CPU, WASM, WebGPU). The merged ONNX file is a single ~52 MB artifact.
import numpy as np
import onnxruntime as ort
session = ort.InferenceSession("smartwatch_lm_merged.onnx", providers=["CPUExecutionProvider"])
x = np.zeros((1, 8), dtype=np.int64)
logits = session.run(None, {"input_ids": x})[0]
print(logits.shape) # (1, 8, 3524)
Quick chat with cleanup:
pip install numpy onnxruntime tokenizers
python onnx_sample.py "How many steps today?"
PyTorch chat
pip install torch tokenizers
python chat.py
from chat import ChatSession
bot = ChatSession(temperature=0.5, max_new_tokens=40, top_k=40)
print(bot.say("How many steps have I taken today?"))
Prompt format
user: How many steps have I taken today?
bot: <INTENT:GET_STEPS> You're at <STEPS_TODAY> of <STEP_GOAL> — keep going!
The model predicts the next token. Slot placeholders like <STEPS_TODAY> are filled by your runtime with live sensor data.
Documentation
| Guide | Description |
|---|---|
| Avoiding gibberish | Special characters to strip, truncation rules, sample scripts |
| Intent reference | All 35 intents and slot placeholders |
| Smartwatch integration | End-to-end device wiring |
Intended use
On-device smartwatch / wearable assistant demos. Not intended for general-purpose chat or safety-critical applications.
Limitations
- Trained on synthetic data; behavior on real user phrasing may vary
- Does not output real metric values — only intents and slot tokens
- Small 12M model with limited reasoning capability