SignalFit-SLM

SignalFit-SLM is a small fitness assistant model for wearable data.

It takes a structured snapshot of a user's recovery, HRV, sleep, training load, recent workouts, goals, and safety flags, then answers a fitness question using only the data it was given. The goal is simple: make wearable coaching more grounded, portable, and safe enough to run locally.

The current model is based on Qwen3-1.7B, fine-tuned with LoRA, wrapped with deterministic safety and grounding checks, and exported as a 4-bit MLX build.

Quick Start

The deterministic tooling and tests require Python 3.10 or newer. Model inference and training require an Apple Silicon Mac.

git clone https://github.com/adidshaft/signalfit-slm.git
cd signalfit-slm
python3 -m venv .venv
.venv/bin/pip install -r requirements-dev.txt
make check

Install MLX support separately when you need to run the model:

.venv/bin/pip install -r requirements-mlx.txt

Example Conversations

Training guidance	Sleep insight	Sleep detail

What It Can Do

• Answer questions like "should I train hard today?", "why is my recovery low?", or "how is my sleep trending?"
• Cite only numbers present in the input context.
• Refuse or redirect unsafe requests instead of coaching through red flags.
• Work across providers once their exports are mapped into the shared schema.
• Run locally through MLX today, with mobile and other runtimes planned.

Current State

SignalFit-SLM has a promoted model of record and a verified 4-bit MLX export. The model passed the project safety bar on the frozen evaluation suite:

Check	Result
No coaching during safety triage	18/18
No protocol details in refusals	19/19
Correct field binding	196/200
Overall deterministic pass rate	135/200

The remaining known failures are quality edge cases around wording and arithmetic, not safety regressions. The promotion rationale is documented in docs/PROMOTION_DECISION_ft_v10.md.

How To Use It Today

Right now the supported path is MLX on Apple Silicon.

1. Create or edit a context JSON.

   Start with examples/sample_context.json. The important part is allowed_numbers: every number the model is allowed to mention must appear there.

2. Run the checked answer wrapper.

   .venv/bin/python scripts/answer_with_check.py \
     --context examples/sample_context.json \
     --expected-action answer_with_caveat \
     --model data/checks/ship-ft_v10/export-4bit \
     -o /tmp/signalfit-answer.jsonl

3. Print the answer.

   python3 -c "import json; print(json.loads(open('/tmp/signalfit-answer.jsonl').readline())['answer'])"

The wrapper is part of the product. It checks model drafts against the same grounding and safety gates used in evaluation, then retries when a draft cites unsupported numbers or violates a safety rule.

Can I Use It In My App?

Yes, but the integration story is still early.

Today, the practical path is:

Target	Status	What to use
Apple Silicon dev machine	Ready	MLX + scripts/answer_with_check.py
iPhone / iOS app	Planned	MLX Swift or Core ML export path
llama.cpp / Ollama-style apps	Planned	GGUF export needed
Web API service	Planned	Thin server around the checked wrapper
Other wearable apps	Ready at schema level	Map data into schemas/assistant_context.schema.json

So if you are building with this today, treat the repo as a working reference implementation: schema, model artifact, wrapper, tests, and evaluation harness. The easiest next step is to point an agent at this repo and ask it to wire the MLX checked-wrapper path into your app.

Roadmap Issues

The open GitHub issues track the missing packaging work:

• MLX Swift / iOS sample integration
• Core ML conversion research
• GGUF / llama.cpp export
• Local HTTP API wrapper
• Provider adapter examples for real wearable exports

Data Model

The model does not connect to wearable accounts directly. Your app or adapter turns provider-specific exports into the shared SignalFit context schema:

Wearable export -> provider adapter -> SignalFit context JSON -> checked model answer

Useful files:

Path	Purpose
schemas/assistant_context.schema.json	Input context schema
examples/sample_context.json	Editable sample context
docs/schema_design.md	Schema design notes
docs/safety_policy.md	Safety policy
docs/testing_guide.md	How to test the current model
docs/process_guide.md	Full training and evaluation history

Repository Layout

Path	Purpose
docs/	Product notes, safety policy, eval plan, process log, promotion notes
benchmarks/	Reproducible human-style benchmark inputs, outputs, and reports
schemas/	JSON Schemas for context and training examples
prompts/	Dataset generation, critique, and evaluation prompts
data/synthetic/	Synthetic training data
eval/v1/	Frozen evaluation suite
scripts/	Validation, serving wrapper, gates, and dataset tools
tests/	Deterministic unit and integration tests
training/configs/	MLX LoRA training configs
data/checks/ship-ft_v10/export-4bit/	Current 4-bit MLX export metadata

Large model weight files are intentionally not committed. Use release artifacts or a model host for deployable weights.

Safety

SignalFit-SLM is not a medical device. It should support fitness decisions, not diagnose conditions or replace professional care.

The assistant must stop coaching and redirect when a request includes medical red flags, unsafe performance-enhancing drug requests, or other high-risk content. See docs/safety_policy.md.

Contributing

Contributions are welcome. Start with CONTRIBUTING.md and the docs/ index. Please use synthetic data only, run make check, and preserve the frozen-evaluation and safety contracts.

• Report vulnerabilities privately through SECURITY.md.
• Community expectations are in CODE_OF_CONDUCT.md.
• Usage and issue-routing help is in SUPPORT.md.

License

SignalFit-SLM is licensed under the Apache License 2.0. The Qwen3 base model is also Apache-2.0; downstream users remain responsible for the licenses of any external datasets, adapters, or export formats they add.

Maintainer

Built by @adidshaft.