---
license: mit
tags:
  - onnx
  - smartwatch
  - intent-classification
  - text-generation
  - wearable
library_name: onnxruntime
pipeline_tag: text-generation
---

# 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 (~28M params) |
| Context length | 256 tokens |
| Vocab size | 3524 (BPE) |
| Training data | Synthetic smartwatch conversations (`tinydata.txt`) |
| Export version | 0.1 |

## Files in this repo

| File | Purpose |
|------|---------|
| `smartwatch_lm_merged.onnx` | Browser/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 chat helper |
| `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 suitable for browser deployment.

```python
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)
```

## PyTorch chat

```bash
pip install torch tokenizers
python chat.py
```

Or in Python:

```python
from pathlib import Path
from chat import ChatSession

bot = ChatSession(
    checkpoint_path=Path("checkpoint.pt"),
    tokenizer_path=Path("tokenizer.json"),
)
print(bot.say("How many steps have I taken today?"))
print(bot.say("thanks"))
```

## 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.

## Intended use

On-device smartwatch / wearable assistant demos. Not intended for general-purpose chat or safety-critical applications.

## Documentation

| Guide | Description |
|-------|-------------|
| [Intent reference](docs/intent-reference.md) | All 35 intents and slot placeholders |
| [Web output quality](docs/web-output-quality.md) | Sampling, cleanup, and anti-gibberish techniques for browser/on-device runtimes |
| [Smartwatch integration](docs/smartwatch-integration.md) | End-to-end guide to wire the model into wrist hardware |

Live browser demo: see the `collab-run-1/web` app in the electron-v1 training project (links to this model on Hugging Face).

## Limitations

- Trained on synthetic data; behavior on real user phrasing may vary
- Does not output real metric values — only intents and slot tokens
- Small model with limited reasoning capability