szl-triage-qwen3.5-0.8b-lora-study5

<!-- SZL-FRONTIER-CRED-20261104 --> <!-- SZL-ESTATE-CARD:v2:START --> <p align="center"><a href="https://a-11-oy.com/"><img src="https://huggingface.co/spaces/SZLHOLDINGS/README/resolve/main/assets/estate-banner-v2.svg" alt="SZL Holdings — governed, receipted, verifiable" width="100%"></a></p>

Control before action. Evidence after.

Product · Proof · GitHub · HF org

<!-- SZL-ESTATE-CARD:v2:END --> <p><a href="https://huggingface.co/spaces/SZLHOLDINGS/szl-command-lab"><img src="https://raw.githubusercontent.com/szl-holdings/szl-brand/27d12e77fc0f9f1a0e5fae8b34adf2380d51d070/kit/logos/orbit-v2/suite/szl-icon-32.svg" alt="SZL Holdings" width="40" height="40" /></a></p>

TypeSafe Triage · Five-Seed LoRA Study

Inspect five LoRA adapters and the frozen training, evaluation and failure evidence behind this structured-triage study.

Artifact: Five PEFT LoRA adapters and retained research evidence · Stage: Research · BLOCKED — 11/12 · NOT_PROMOTABLE

Explore in Command Lab · Build · Evidence

Before you use it

• The recorded contamination verdict remains a release blocker. Published adapters do not establish deployment clearance.
• Results describe the frozen 113-row family holdout; they do not establish generalization, calibration or production readiness.
• The earlier 66-row gate report does not clear this later five-seed study.
• The inline inference example remains withdrawn. Read the original implementation caveat below before planning an experiment; this card supplies no new runtime evidence.

<details> <summary>Technical details and original evidence</summary>

The complete canonical source body follows, including all measured results, historical gate scope, contamination limits and the withdrawn inference example notice.

<!-- SZL-PRESERVED-TECHNICAL-BODY:START -->

SZL TypeSafe Triage · Five-Seed LoRA Study

A measured model artifact with its limits attached.

This repository contains five LoRA training runs for structured triage: seeds 11, 23, 37, 53, and 71. It includes the default adapter at the repository root, every seed under adapters/, frozen evaluation identity, raw predictions, failure records, training receipts, and aggregate metrics.

Status

Published does not mean promoted. The owner authorized publication of the trained research artifact. The existing contamination verdict and release boundary remain visible instead of being removed.

Historical gate scope

The root gate_report.json records an earlier 66-row gate with verdict PROMOTABLE. It is retained historical evidence, not promotion of this later five-seed, 113-row study. The current study remains BLOCKED — 11/12 and NOT_PROMOTABLE. Published adapter files and historical gate labels do not supersede this release boundary.

What it does

The adapter accepts a triage input and is trained to return only:

{"label":"...","state":"...","evidence":["..."]}

In plain language:

• label is the model's proposed category.
• state records whether it can decide or needs review.
• evidence contains text spans intended to come directly from the input.
• A REVIEW state is a refusal to overclaim.

Study design

• Base model: unsloth/Qwen3.5-0.8B
• LoRA rank: 16
• LoRA alpha: 16
• Epochs: 3
• Learning rate: 0.0002
• Precision: bfloat16
• Quantization during training: none
• Effective batch size: 4
• Training rows: 515
• Held-family rows: 113
• Seeds: 11, 23, 37, 53, 71
• Split rule: template families are kept together
• Evaluation decoding: greedy
• Malformed JSON repair: disabled

Measured results

Across the five adapters, mean joint accuracy was 100.0%, with sample standard deviation 0.0%.

These results describe this frozen 113-row family holdout. They do not establish broad generalization, calibration, or production readiness.

Evidence grounding checks whether each generated span occurs in its input; exact target evidence compares the complete ordered evidence list with the retained target list. These are different measures, with separate denominators shown in the table: generated spans for grounding and held rows for exact evidence. The evaluator's failure rows cover malformed JSON, joint label/state mismatch, or ungrounded spans; they do not include an exact-evidence-only mismatch. Thus zero failure rows does not establish perfect exact target evidence. The base's exact-evidence count includes empty-list comparisons despite zero valid JSON rows; it must not be interpreted as valid base predictions.

These are retained historical results. This card-only correction performs no new model evaluation, model loading, training, release clearance, or publication.

Repository layout

README.md
adapter_config.json
adapter_model.safetensors
tokenizer files
adapters/
  seed-011/
  seed-023/
  seed-037/
  seed-053/
  seed-071/
evidence/
  frozen/
  training/
  evaluation/
  adapter-index.json
  environment-receipt.json

The root adapter is seed 11, retained as the tagged first measured run. The additional seed directories support reproducibility and stability inspection.

Inference implementation

Inline inference example withdrawn. The previous generated quick start contained malformed output indexing and could also route text through the wrong processor interface. No fresh inference was performed for this card correction.

Inspect the canonical evaluation implementation alongside the retained environment receipt and adapter identities before a new experiment. The historical script's main entry point combines training, evaluation, and publication; it is not a card-only repair command or a standalone inference quick start. This documentation change supplies no new runtime, held-out, or promotion evidence.

Using another seed

Download the repository and load one of these local directories:

adapters/seed-023
adapters/seed-037
adapters/seed-053
adapters/seed-071

Each directory is a complete PEFT adapter. The SHA-256 digest for every adapter is recorded in evidence/adapter-index.json.

Evaluation rules

The held set was separated by template family rather than random row. This reduces direct template leakage between training and evaluation.

Every target was scored for:

• Strict JSON validity
• Exact label agreement
• Exact state agreement
• Joint label-and-state agreement
• Refusal fidelity
• False labels on refusal cases
• Evidence grounding
• Exact target-evidence agreement

Raw predictions and complete failure records are included so the headline numbers can be audited.

Limitations

• datasets YAML undeclared: no Hub datasets: id is asserted on this card. Local/receipt digests (jsonl SHA) are not Hub dataset ids; this wave does not invent them.

• Release remains blocked at 11/12.

• Promotion has not been established.

• The recorded contamination verdict still travels with the artifact.

• The model is not calibrated.

• The study does not establish generalization outside the frozen holdout.

• The model is not a replacement for the deterministic decision engine.

• Generated output must be parsed and validated before downstream use.

• High-impact decisions require human or deterministic review.

Reproducibility

Source repository:

https://github.com/szl-holdings/szl-typesafe-triage

First measured run:

https://github.com/szl-holdings/szl-typesafe-triage/releases/tag/triage-lora-run1

Five-seed evidence tag:

https://github.com/szl-holdings/szl-typesafe-triage/releases/tag/triage-lora-study5-measured-20260922-111314

Citation

If this artifact is discussed, describe it as:

SZL TypeSafe Triage five-seed LoRA study: training and frozen evaluation measured; public experimental artifact; promotion not established.

<!-- SZL-PRESERVED-TECHNICAL-BODY:END --> </details>

Artwork: SZL Orbit · © 2026 SZL Holdings / Stephen Lutar · CC BY 4.0.

The estate

Provenance note: duplicate weights

The listed weight files in this repository are byte-identical (same LFS SHA-256; checkpoint-* excluded) to files in another repository: SZLHOLDINGS/szl-triage-qwen3.5-0.8b-lora-study5-laptop-lora@f5524ca958c4 (shared: adapter_model.safetensors, adapters/seed-011/adapter_model.safetensors) (checked 2026-11-12 at 2e893ecef493). Identical bytes are one artifact published under more than one name; this repository is not independent evidence of a separate training run, and any results stated for it apply to that same artifact. Existing result lines above are left unchanged.