# Unified 1-1-1 Hugging Face release contract

## Repository identity

- New repository: `mp-juuuns/qwen35-standalone4l-classification-base`
- Existing repository preserved without mutation:
  `mp-juuuns/qwen35-standalone4l-propaganda-classifier`
- Existing specialized fixed revision:
  `e7e0ff16828052687d2e8dd7849e7a521629cf38`

## The three public surfaces

1. Repository root: a headless four-layer Qwen3.5 text backbone trained without
   SemEval examples, labels, logits, thresholds, or evidence objectives.
2. `models/semeval-propaganda/`: the existing three-seed task-specific
   classifier copied byte-for-byte as a separate historical lineage. It is not
   described as a fine-tuned child of the root base.
3. `distillation/`: the reusable 24L→8L→6L→4L platform for unlabeled
   representation distillation and user-owned single-label or multilabel data.

## Base training contract

- Source: text backbone and full tokenizer of `Qwen/Qwen3.5-0.8B`.
- Structural defaults:
  - 24→8: `0,4,6,11,13,16,20,23`
  - 8→6: `0,1,3,4,6,7`
  - 6→4: `0,2,3,5`
- Task-free data: 4,096 deterministically shuffled non-empty rows from the
  first locally cached WikiText-103 raw training shard.
- Objective: aligned hidden-boundary and final-representation distillation.
- One epoch per stage, seed 41, maximum length 128, batch size 1, gradient
  accumulation 8, learning rate 2e-5, BF16 CUDA.
- Root artifact: 334,096,704 parameters and 668,198,976 BF16 weight bytes.
- No classification head is published at the root.

WikiText source text is not redistributed. Its dataset page's metadata and
prose currently disagree on the license version, so users must review the
upstream terms directly.

## Benchmark evidence contract

The headless root is evaluated through a fresh-head transfer probe, not by
pretending it is already a classifier. Both the task-agnostic base and the
structural-copy control use the same frozen protocol:

- SemEval-2020 Task 11 annotations converted to article-level 14-label targets;
- train/calibration/test article counts 260/56/55;
- seeds 41/42/43; five epochs; no early stopping;
- 256-token windows, stride 128, label-wise maximum article aggregation;
- micro batch 1, accumulation 32, AdamW, LR 2e-5, BF16;
- one global threshold selected on epoch-5 calibration Macro-F1 then Micro-F1;
- 55 previously opened test articles and 434 test windows.

| Initialization | Macro-F1 mean ± SD | Micro-F1 mean ± SD | Exact match mean ± SD |
|---|---:|---:|---:|
| Task-agnostic KD base | 0.58748 ± 0.01579 | 0.65295 ± 0.01088 | 0.03030 ± 0.02777 |
| Structural copy, no task-free KD | 0.51809 ± 0.01631 | 0.59046 ± 0.01603 | 0.00000 ± 0.00000 |

The observed same-seed mean Macro-F1 difference is +0.06939. All values and
sample SDs must be derived from the six frozen full reports at build time; the
card, JSON, and CSV tables must numerically close against those reports.

The existing specialized 4L lineage may appear only as a clearly labeled
reference because it uses a separate task-specific shrink/distillation history
and reduced vocabulary. It is not evidence that the task-free base produced
that checkpoint.

The report bundle must include:

- per-arm, per-seed Macro-F1, Micro-F1, exact match, positive rate, threshold;
- three-seed means and sample SDs;
- per-label precision, recall, F1, and fixed test support;
- parameters, CUDA allocation/reservation, total wall time, and test timing;
- compression-ladder depth, parameter count, BF16 size, hashes, and KD losses;
- normalized full frozen reports with article-level probability vectors.

The test is already opened, per-label supports are small, and three seeds are
not a confidence interval. No confirmatory, significance, universal
classification, or population-generalization claim is permitted.

## Multilingual metadata contract

The root keeps the full upstream tokenizer. The official Qwen3.5 card states
support for 201 languages and dialects and reports upstream multilingual
benchmarks. Therefore `language: multilingual` and a `multilingual` tag are
permitted as inherited input/architecture metadata.

This release's task-free KD corpus and only downstream probe are English.
Accordingly, the card must state prominently that the four-layer root has no
direct multilingual downstream evaluation and that tokenizer coverage is not a
multilingual classification-quality result. It may not claim verified quality
across 201 languages.

## Release gates

- Root `model.safetensors` contains no `score.weight`.
- Root loads with `AutoModel`; a temporary head can be attached with
  `AutoModelForSequenceClassification`.
- Root weight SHA-256 is
  `2732c616772fe320cdea228ab4554981418b1b2bf615c4183fb1ac8e6e2168d3`.
- Base manifests say `labels_read=false` and `semeval_used=false`.
- Root metadata contains `multilingual` and contains no propaganda task tag.
- Benchmark CSV/JSON/card values close against all frozen reports.
- Nested specialized checkpoint hashes equal the existing release.
- The platform validates label order and ID/group split leakage, never
  overwrites output directories, hashes inputs/outputs, verifies immutable
  teachers, and reloads fresh checkpoints.
- The release contains no SemEval or WikiText source data.
- The old specialized repository revision and inventory are checked before and
  after updating only the new repository.