# NOTICE — MetaFLOS HMI Transcripts

Dataset: **MetaFLOS HMI Transcripts** (Manufacturing Human-Machine Interaction Dialogue Dataset)
Author / rights holder: **Tzer-jen Wei**
License: **CC BY 4.0** (see `LICENSE`)

This file is informational. It records provenance, upstream terms, and
safety advisories. **Nothing here adds conditions to the CC BY 4.0 grant** —
CC BY 4.0 §2(a)(5)(B) forbids downstream restrictions, and none are intended.

---

## 1. Provenance

All 30 scenarios / 325 turns are **synthetic**, produced by `gemma3:27b`
via `scripts/gen_hmi_dataset.py` (seed scenario → LLM transcript → role and
non-empty validation). No real conversation, factory event, or personal data
is contained.

## 2. Gemma (Google) — verified position

Verified against the Gemma Terms of Use (last modified 2026-04-01) and the
Gemma Prohibited Use Policy (last modified 2024-02-21):

- **Gemma Terms §1.1(e)**: "For clarity, Outputs are not deemed Model
  Derivatives."
- **Gemma Terms §3.3**: "Google claims no rights in Outputs you generate
  using Gemma."

This dataset consists solely of Outputs. The redistribution conditions of
**Gemma Terms §3.1** — enforceable pass-through of use restrictions, delivery
of the Gemma Agreement, NOTICE file — attach to Gemma and to Model
Derivatives, and therefore **do not attach to this dataset**. The author
generated it in compliance with the Gemma Prohibited Use Policy, as §3.2
requires of the party operating the model.

**If you fine-tune on this dataset**: Gemma Terms §1.1(e) defines "Model
Derivatives" to include a model created by "methods based on the generation
of synthetic data Outputs by Gemma for training that model" so as to cause it
"to perform similarly to Gemma". A model you distill or fine-tune with that
effect may itself be a Gemma Model Derivative and become subject to the full
Gemma Terms. Training a *different* base model for *manufacturing-domain*
capability is a different case. Assess your own situation before distributing.

- https://ai.google.dev/gemma/terms
- https://ai.google.dev/gemma/prohibited_use_policy

## 3. Scenario inspiration sources

20 of the 30 scenarios **name** a public academic dataset as the origin of the
defect **category** that inspired the scenario (e.g. "surface scratch on a
metal nut, MVTec AD `metal_nut` class").

**No text, image, annotation, or other expressive content from any of these
sources is reproduced, adapted, or redistributed here** — only factual
category and domain references are used. Accordingly, no ShareAlike or other
copyleft obligation of those sources attaches to this dataset.

| Source | Holder | Scenarios |
|---|---|---|
| MVTec AD | MVTec Software GmbH | 6 |
| IRWoZ 1.0 / 2.0 | Aalborg University | 4 |
| NEU-DET | Northeastern University | 3 |
| DeepPCB | Tang et al. | 3 |
| KIDE4I | Tekniker Foundation | 2 |
| ICNLSP 2023 corpus | respective authors | 1 |
| KolektorSDD | Kolektor Group / UL FRI | 1 |
| Self-authored | Tzer-jen Wei | 10 |

If you obtain and use any of these upstream datasets **directly**, you must do
so under that dataset's own license, which this notice neither grants nor
alters.

## 4. Safety advisory — not engineering guidance

The parameter values, setpoints, tolerances, and corrective actions in the
dialogues are **synthetic illustrations**, not validated engineering guidance
for any real machine. They have not been checked against any real process.

Machine identifiers (`NX-220`, `EL-180`, …) and all numeric parameters are
fictional scenario settings and correspond to no real production line,
equipment configuration, or trade secret.

**The author strongly advises** against using this dataset — or any model
trained on it — to drive, recommend, or automate real machine control,
process parameters, maintenance decisions, or safety interlocks without
independent qualification by a responsible engineer.

## 5. Transparency advisory

When quoting or redistributing these dialogues, please identify them as
LLM-generated synthetic text rather than records of real human-machine
interaction. See also EU AI Act transparency obligations for synthetic
content, where applicable to you.

## 6. Attribution

CC BY 4.0 §3(a)(1) requires attribution. Suggested form:

> MetaFLOS HMI Transcripts, by Tzer-jen Wei, licensed under CC BY 4.0.
> https://huggingface.co/datasets/tjw/hmi-transcripts

```bibtex
@misc{metaflos_hmi_transcripts_2026,
  title  = {MetaFLOS HMI Transcripts: A Manufacturing Human-Machine
            Interaction Dialogue Dataset},
  author = {Wei, Tzer-jen},
  year   = {2026},
  note   = {30 scenarios / 325 turns, synthetically generated with gemma3:27b},
  url    = {https://huggingface.co/datasets/tjw/hmi-transcripts}
}
```

---

# 中文說明

本檔案為**告知性質**，記錄資料來源、上游條款與使用建議。
**其內容不對 CC BY 4.0 之授權附加任何條件** —— CC BY 4.0 §2(a)(5)(B) 禁止
下游附加限制，本文件亦無此意圖。

**1. 資料來源**：30 情境／325 輪全為 `gemma3:27b` 生成之合成內容，無真實
對話、真實工廠事件或個人資料。

**2. Gemma**：依 Gemma 使用條款 §1.1(e)「Outputs 不視為 Model Derivatives」
及 §3.3「Google 不主張 Outputs 之權利」，本資料集全部為 Outputs，故 §3.1 之
散布條件不及於本資料集。著作人已依 §3.2 於生成時遵守 Gemma 禁止用途政策。
**惟若您以本資料集微調出表現近似 Gemma 之模型**，該模型本身可能構成 Model
Derivative 而受完整 Gemma 條款拘束；若是在其他 base model 上微調製造領域能力，
則屬另一情形。請於散布前自行評估。

**3. 情境靈感來源**：30 個情境中有 20 個標註了啟發該情境之瑕疵**類別**所出自
的公開學術資料集。**未重製或改作任何來源之文字、影像或標註**，僅使用事實性
類別指涉，故不生 ShareAlike 義務。來源對照見上方英文表格。若您**直接取用**
該等來源，仍須依其各自授權。

**4. 安全提醒**：對話中的參數值為**合成示例**，非經驗證之工程指引；機台代號與
數值均為虛構情境設定。著作人**強烈建議**：未經負責工程師獨立驗證，不要將本
資料集或以其訓練之模型用於驅動真實機台控制、製程參數、維修決策或安全連鎖。

**5. 透明性提醒**：引用或再散布時，請敘明其為 LLM 生成之合成文本。

**6. 姓名標示**：CC BY 4.0 §3(a)(1) 要求標示。建議格式見上方英文段落。