prathamkode/smartwatch-lm-0.2

Smartwatch LM v0.2

Small GPT for wrist-wearable chat. The model replies in natural language and emits intent tags like <INTENT:GET_STEPS> plus slot placeholders such as <STEPS_TODAY> that your app fills from live sensor data.

Model details

Property	Value
Architecture	6-layer causal GPT (~15.4M params)
Context length	256 tokens
Vocab size	5533 (BPE)
Best val loss	0.3243
ONNX size	~60 MB
Export version	0.2

Files

File	Purpose
smartwatch_lm_merged.onnx	On-device inference (ONNX Runtime, opset 17)
checkpoint.pt	PyTorch weights
tokenizer.json	BPE tokenizer
config.json	Architecture + ONNX I/O
model.py / chat.py	PyTorch load + REPL
reply_utils.py	BPE cleanup, intent parse, slot fill
onnx_sample.py	ONNX generate sample
docs/	Integration, intent reference, output cleanup
benchmark/	Golden prompts, quality report, charts

Quick start

pip install numpy onnxruntime tokenizers
python onnx_sample.py "How many steps today?"

pip install torch tokenizers
python chat.py

ONNX I/O

• Input: input_ids int64 [batch, seq] (max seq = 256)
• Output: logits float [batch, seq, vocab_size]

Sample logits at the last position autoregressively until EOS or max_new_tokens. Recommended settings: temperature=0.5, top_k=40, max_new_tokens=40.

Documentation

Guide	Description
Avoiding gibberish	BPE cleanup, truncation rules, sample scripts
Intent reference	All 35 intents and slot placeholders
Smartwatch integration	End-to-end device wiring

Benchmarks

Quality evaluation on 39 golden prompts is in benchmark/:

• benchmark/report.json — full per-prompt results
• benchmark/charts/ — metric charts
• benchmark/benchmark_prompts.json — golden prompt set

License

This model is released under the MIT License.