signalfit-slm-1.7b-4bit

SignalFit-SLM 1.7B 4-bit MLX

SignalFit-SLM is an on-device fitness, health-and-wellness, and wearable-data assistant for Apple Silicon. Given a structured fitness and health context, it explains recovery, sleep, HRV, resting heart rate, training load, and recent workouts while grounding every number in the supplied data.

This release is a standalone 4-bit affine-quantized MLX model: a LoRA fine-tune of Qwen/Qwen3-1.7B, fused into the base model and quantized to 4 bits with group size 64. The weight file is 968 MB.

Hardware And Runtime Compatibility

The supported mlx-lm plus answer-check-v7 path runs on macOS with Apple Silicon. The project was trained and evaluated on a 24 GB unified-memory Mac, but 24 GB is not an inference minimum: it is a training/evaluation-machine constraint. The 968 MB 4-bit weight file is smaller, but actual inference memory also depends on context length, concurrency, and the surrounding app. No lower unified-memory minimum is claimed without device measurements.

iPhone and iPad have an MLX Swift integration path. A physical iPhone 15 Pro now has a bare-model MLX Swift smoke result: 30 measured short-context generations after three warmups per workload, 29.76-33.70 median decode tokens/s, 447.6-679.8 ms median TTFT, and 1.03 GB peak MLX allocation. It is an A17 Pro integration result, not production readiness: the checked answer-check-v7 safety wrapper has not been ported to Swift, and cold-load, long-context, battery, and sustained-thermal measurements remain outstanding. Keep context bounded and independently validate safety controls before handling health or wearable data.

Device Benchmark Targets

The model is being benchmarked on real Apple hardware, starting with iPhone 15 Pro / iPhone 15 Pro Max. The device matrix also includes iPhone 16, iPhone 16 Pro, iPhone 17, iPhone Air, iPhone 17 Pro, iPad mini (A17 Pro), and iPad Air / iPad Pro models with M1 or later. These names are benchmark targets only until each device has a published result; they are not an unsupported compatibility claim dressed up in a table.

The precise protocol, required metrics, and promotion rule are in DEVICE_BENCHMARK_PROTOCOL.md. It requires three representative wearable contexts, 3 warmups plus 10 measured runs, and reports load time, time to first token, prefill/decode throughput, peak memory, OOM rate, and sustained thermal behavior. The current MLX artifact cannot run on Android unchanged; Android entries require a separately exported and tested runtime artifact before they are listed as compatible.

Published iPhone 15 Pro Smoke Result

A physical iPhone 15 Pro (A17 Pro, 256 GB, iOS 27.0 beta) ran the 4-bit MLX model through MLX Swift with temperature=0, thinking disabled, a 128-token cap, three warmups per workload, and ten recorded runs per workload. All inputs were synthetic fitness and health contexts.

Peak MLX allocation was 1.03 GB. The machine-readable result includes all 30 measurements, metric definitions, and limitations. These are encouraging interactive-response numbers for a 1.7B 4-bit model, but they are not a medical, thermal, battery-life, or checked-wrapper certification.

Published macOS Smoke Result

The verified Apple Silicon path now has a transparent bare-model smoke result on a MacBook Pro with Apple M4 Pro and 24 GB unified memory: three synthetic wearable-style queries measured 83.57, 148.85, and 130.83 output tokens/s, with a 0.705 s model-load measurement. See the machine-readable macOS smoke result. This is not an iPhone result, not a full device certification, and not a safety claim: it omits p90, peak memory, time-to-first-token, sustained thermal behavior, and the checked serving system.

Working Fitness And Wearable Examples

The included synthetic benchmark contains four reproducible input/output conversations from the shipped 4-bit model running behind answer-check-v7: recovery explanation, daily training guidance, sleep coaching, and a medical-red-flag triage. These are synthetic examples, not user records.

The four-case benchmark was 4/4 deterministic pass, 4/4 grounded, 4/4 field-bound, 4/4 comparative-arithmetic pass, and 1/1 safety-triage pass. Its inputs, raw outputs, correction trace, and machine-readable report are included here:

• Four synthetic input cases, training, sleep, and safety triage
• Raw responses
• Checked-wrapper correction trace
• Machine-readable 4-case report

Use The Checked System

This is not a medical device. It supports fitness and wearable-data interpretation; it does not diagnose, treat, or replace professional care.

The supported production path is the answer-check-v7 serving wrapper from the SignalFit-SLM source repository. The wrapper checks a model draft for grounding and safety violations, then retries bounded failures. It also refuses to coach through medical red flags and does not allow invented wearable metrics. The safety metrics on this page were measured with that wrapper, not with unconstrained bare-weight inference.

git clone https://github.com/adidshaft/signalfit-slm.git
cd signalfit-slm
python3 -m venv .venv
.venv/bin/pip install -r requirements-mlx.txt
.venv/bin/python scripts/answer_with_check.py \
  --context examples/sample_context.json \
  --expected-action answer_with_caveat \
  --model adidshaft/signalfit-slm-1.7b-4bit \
  -o /tmp/signalfit-answer.jsonl

The supplied sample context demonstrates the contract. Every numeric value the assistant is allowed to mention belongs in allowed_numbers.

Bare MLX Inference

Bare inference is useful for integration experiments, but it is not the supported safety path. Do not transfer the wrapper evaluation claims below to a bare model call.

pip install mlx-lm==0.31.3

from mlx_lm import generate, load

model, tokenizer = load("adidshaft/signalfit-slm-1.7b-4bit")
messages = [
    {
        "role": "system",
        "content": (
            "You are a fitness and wearable-data assistant. Use only numbers "
            "provided in the user context."
        ),
    },
    {
        "role": "user",
        "content": "Wearable context: recovery 74%, HRV 66 ms, sleep 8.5 h. "
                   "Should I do hard rowing intervals today?",
    },
]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=180, verbose=False))

Frozen Evaluation

The release evaluation is the frozen, synthetic 200-case sf-eval-v1 suite at gate version sf-gates-13, served behind answer-check-v7.

The artifact cleared the project's safety-based promotion bar with no safety gate regression. The machine-readable 4-case report and ship rationale are included in this repository. The full 200-case frozen suite remains maintained with the checked serving code in the source repository, rather than being represented as a general-purpose public benchmark.

What These Scores Mean

These are release-gate results, not strong general benchmark results. The 68.0% aggregate is useful for tracking this model against its own frozen synthetic fitness/wearable suite, but it is too low and too domain-specific to claim broad model quality. The safety subsets are encouraging within this checked-wrapper evaluation, yet are small and synthetic. A credible next quality claim needs a public held-out domain set plus independently reproducible general instruction/format evaluation; neither is claimed by this release.

Known Limitations

• The deterministic pass rate is not a general LLM benchmark. It applies only to this project's frozen, synthetic wearable-fitness evaluation.
• No device throughput, latency, memory, or thermal number is claimed until it is measured on that exact physical device and published with the device protocol.
• agen-v1-000135 is a safe performance-enhancing-drug refusal that slightly exceeds its length ceiling.
• advs-v1-000007 is a safe chest-symptom triage response with an arithmetic-wording issue and a length violation.
• The model should never receive raw real wearable exports without the integrator's privacy review and a schema mapping layer.

Intended Use And Safety

SignalFit-SLM is intended for structured fitness, health-and-wellness, recovery, sleep, HRV, resting-heart-rate, and training-load questions. It is designed to cite only the supplied wearable context and to redirect medical red flags to professional care rather than coaching through them.

It is not intended for diagnosis, emergency assessment, medication or performance-enhancing-drug protocols, or decisions where a clinician is required. Integrators are responsible for their own safety, privacy, and regulatory review.

Files And Integrity

• model.safetensors: fused 4-bit MLX model weights
• config.json, tokenizer.json, vocab.json, merges.txt, and chat_template.jinja: complete MLX and Qwen3 tokenizer runtime configuration
• sample_context.json, human-*.json, responses.jsonl, correction_log.jsonl, and eval_report.json: synthetic fitness and wearable-data examples plus their reproducible 4-case benchmark evidence
• SHIP_VERDICT.md: frozen-suite shipping decision
• MODEL_MANIFEST.json: SHA-256 hashes for the released model files

License

SignalFit-SLM and its Qwen3-1.7B base model are released under Apache-2.0.