M1 auth rule specialist — reference seed 17
Author: Megern Qaisse (megern). Source, reproducible protocols and all seeds: M1 Memory Research.
An original local LoRA adapter over mlx-community/Qwen3-0.6B-4bit, trained on an Apple M1 with 16 GiB unified memory. This contains adapter weights only. Base revision: 73e3e38d981303bc594367cd910ea6eb48349da8.
The task is authentication-event priority classification from three Boolean policy features computed by Python. Prompts state the rule. No real incident logs, hosted training or teacher API was used. These synthetic tasks can also be solved directly by rule-based software.
Recorded evaluation
Final seed-17 checkpoint on the frozen 96-case synthetic test: strict accuracy 96/96 (100.00%), macro-F1 1.0000, valid JSON 96/96. All three training seeds and a common base baseline must be read together in the source study; this adapter was not selected for its test score.
The same task's unadapted base classification accuracy is 48.96% after removing only a complete outer Markdown JSON fence. Base strict schema accuracy is reported separately. This keeps formatting failures distinct from wrong labels. This exploratory prototype uses different sampling, language balance, prompts and feature representation. Python performs the count and time comparisons; the LLM does not. There are only eight possible Boolean states. A direct rule baseline solves the entire closed task; model scores do not demonstrate numerical reasoning or general security competence.
Local use
Install the pinned study dependencies and acquire the pinned base checkpoint locally. Then use the study evaluator with --model /absolute/local/base --adapter /absolute/local/this-adapter --data /absolute/local/test.jsonl --output /absolute/local/fresh-report.json.
Generation uses the base chat template with enable_thinking=False, greedy decoding and a 48-token output cap. The adapter expects the study's system instruction, a stated classification rule, and JSON evidence. It is not a general conversational model.
Training and limitations
384 training and 48 validation examples, 400 iterations, batch one, last 16 layers, query/value LoRA rank eight, learning rate 0.0002, prompt masking and gradient checkpointing. No test-based checkpoint selection. See provenance.json for hashes, measured allocation and timing. MLX allocation is different from total process RAM.
This is an experimental synthetic classifier. Scores do not demonstrate general Arabic proficiency, production SOC accuracy, adversarial robustness or security guarantees. No peer review or novel training algorithm is claimed. Base weights remain under their Apache-2.0 terms; this adapter is released under Apache-2.0. The original study tools and synthetic data use MIT.
Use tools/predict.py --normalize --task auth with the study base, this adapter and a raw evidence record. Omitting --normalize changes the evaluated interface.
The complete same-budget series contains seed accuracies 100%, 100% and 66.67% (mean 88.89%). Seed 41 fails every medium case; this reference adapter is seed 17 by the protocol, not a best-score release. Direct rule code scores 100%.
Quantized