plawanrath/static-student-tel-mlx-S-w4

A 4-bit quantization-aware byte-level telemetry student (S: 4 layers, d_model 192, 4 heads), distilled from the float student plawanrath/static-student-tel-mlx-S-fp, from the static-student project (Weights as Constants: compiling distilled task models into native executables as read-only data). Code, RUNBOOK and results: https://github.com/plawanrath/static-student.

What the model does

Input: the raw bytes of one telemetry payload (a metric line in one of many ad-hoc formats). Output: per-byte span pointers for the metric name, value and unit, a unit class, a value score, and a defer logit. The student is meant to run only on payloads a frozen regex parser rejects, and to answer only when its confidence clears a threshold certified on the integer kernel that ships; everything else is deferred. It is a research artifact for one synthetic task family, not a general log parser.

Training

curriculum	data/curriculum/telemetry-mlx-s0-x12k (written by Mistral Small 3.2 24B from the spec line, expanded with drift rewrites)
init / distillation teacher	models/tel-mlx-S-fp / models/tel-mlx-S-fp
rendered training rows	500,000
epochs	2
seed	0
parameters	1,936,330
packed weight bytes (trunk / total at 4 bit)	1,022,976 / 1,127,626
dev struct-exact after quantization, before fine-tuning	99.09%
dev struct-exact, final (step sizes snapped to float16, as shipped)	99.36%
wall time (Apple silicon, MPS)	31 min

Certificate of the compiled kernel

The weights were compiled into a C kernel with the weights as a constant array; the deferral threshold below was certified (Learn-then-Test) on that exact integer kernel, on a locked calibration pool, and travels in the binary.

risk bound α / confidence δ / min coverage	0.01 / 0.05 / 0.5
calibration pool n	20,000
threshold τ	-0.143815
certified coverage	78.86% (15,771 accepted, 86 wrong)
empirical risk on accepted	0.55%
thresholds tested	96
pool sha256	e06a4a53395c4486…

Files

• model.safetensors, config.json — framework-neutral weights and constructor arguments.
• student.pt — the original PyTorch checkpoint the RUNBOOK commands expect (torch.load, dict with config, state, bits).
• train_report.json — the training log above.

How to load

git clone https://github.com/plawanrath/static-student && cd static-student && bash scripts/env/setup.sh

from static_student.student.release import load_released
model, cfg = load_released("plawanrath/static-student-tel-mlx-S-w4")

To compile it into a certified binary (needs a C compiler, make, RE2):

.venv/bin/python scripts/release/download_hf.py tel-mlx-S-w4
.venv/bin/python -m static_student.build --spec specs/telemetry.spec --model models/tel-mlx-S-w4 --out build/tel-mlx-S-w4

Limitations

• Trained and evaluated on synthetic telemetry payloads rendered from format families; never on production traffic. Numbers above are on held-out synthetic families.
• The certificate is a statistical bound on the rate of wrong accepted answers under the calibration distribution; it says nothing about inputs from other distributions.
• The research project was concluded without a publication; the model is released as-is for reuse and reproduction and is not maintained.

Citation

If you use this checkpoint, please cite the repository:

@software{rath2026staticstudent,
  author  = {Rath, Plawan Kumar},
  title   = {static-student: Weights as Constants --- compiling distilled task models into executables as read-only data},
  year    = {2026},
  doi     = {10.5281/zenodo.23202887},
  url     = {https://github.com/plawanrath/static-student},
  license = {Apache-2.0},
  version = {0.1.0}
}

CITATION.cff in the repository carries the same entry (GitHub's "Cite this repository" button).

Disclaimer

Plawan Kumar Rath (@plawanrath). This work was conducted in the author's personal capacity. The views expressed here are those of the author and do not necessarily reflect the views of Meta.