Datasets:

tjw
/

hmi-transcripts

license: cc-by-4.0
language:
  - en
  - zh
language_bcp47:
  - zh-TW
tags:
  - manufacturing
  - human-machine-interaction
  - dialogue
  - vicuna-style
task_categories:
  - text-generation

MetaFLOS HMI Transcripts (Manufacturing Human-Machine Interaction Dialogue Dataset)

Multi-turn dialogue dataset between manufacturing floor operators and machine intelligent assistants, generated in Vicuna style (seed scenario → LLM produces transcript).

Content

30 scenarios / 325 dialogue turns (6–12 turns per scenario)
Covers 16 domains: textile warping/weaving, LCD panels, semiconductor processes, fans/blowers, water chillers, general equipment reliability (FMEA), robot collaboration, maintenance & repair, technical document support, steel/metal, PCB, metal/fasteners, electronic components, building materials, cables, wood processing
Dialogue arc: detect defect → diagnose cause (citing defect database) → adjust parameters (concrete values) → verify result (measurement/re-inspection)

Sources

Source	# Scenarios
Self-built scenarios (Vicuna-style generation)	10
IRWOZ 2.0 / IRWoZ 1.0 (Aalborg Univ.)	4
KIDE4I (Tekniker Foundation, Industry 5.0)	2
ICNLSP 2023 industrial technical document corpus	1
NEU-DET (hot-rolled steel strip surface defects)	3
DeepPCB (PCB surface defects)	3
MVTec AD (general industrial anomalies)	6
KolektorSDD (metal surface micro-cracks)	1

Schema (JSONL, one record per line)

{
  "id": "hmi-18",
  "domain": "Steel/Metal",
  "machine": "Hot-rolled steel strip line",
  "scenario": "Scenario description",
  "source": "NEU-DET (hot-rolled steel strip surface defects)",
  "transcript": [
    {"role": "operator", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "model": "gemma3:27b",
  "ts": "2026-08-09 19:00:00"
}

Generation Method

scripts/gen_hmi_dataset.py: seed scenario → ollama (gemma3:27b) generates a 6–12 turn transcript, validates roles are legal and content non-empty.

Privacy Protection

Synthetic content: all scenarios and dialogues are LLM-generated, not real factory events or real personnel conversations.
No personal data: transcripts contain no real names, phone numbers, emails, addresses, or other personally identifiable information; operators are always presented as the "operator" role.
No sensitive machine data: parameter values in dialogues are synthetic scenario settings, not actual parameters of any real production line.

License & Copyright

This dataset: CC BY 4.0. Free to use, share, and adapt — including commercially — with attribution. CC 4.0 §4 also covers sui generis database rights.
Notices: see NOTICE.md for provenance, upstream terms, and safety advisories. Those are informational and add no conditions to the CC BY 4.0 grant.
Scenario sources: 20 of 30 scenarios name a public academic dataset as the origin of the defect category that inspired them (e.g. "MVTec AD metal_nut class"). No source text, image, or annotation is reproduced or adapted — only factual category references — so no ShareAlike obligation attaches. Credits in NOTICE.md.
Generation model: gemma3:27b. Gemma Terms §1.1(e)/§3.3 place Outputs outside the §3.1 redistribution conditions. If you fine-tune to Gemma-like behaviour, your model may itself become a Gemma Model Derivative — see NOTICE.md §2.
⚠️ Not engineering guidance: dialogue parameter values are synthetic illustrations. Do not drive real machine control or safety interlocks without independent qualification by a responsible engineer.

Author & Citation

Tzer-jen Wei

@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

MetaFLOS HMI Transcripts（製造現場人機互動對話資料集）

製造現場操作員 ↔ 機台智慧助理的多輪對話資料集，Vicuna 式生成（seed 情境 → LLM 產生 transcript）。

內容

30 個情境 / 325 輪對話（每情境 6~12 輪）
涵蓋 16 個領域：紡織整經/織造、面板 LCD、半導體製程、風機/送風、水機/冰水主機、通用設備可靠度(FMEA)、機器人協作、維修保養、技術文件支援、鋼鐵/金屬、PCB、金屬/扣件、電子元件、建材、線纜、木材加工
對話弧線：發現瑕疵 → 診斷成因（引用瑕疵資料庫）→ 調校參數（具體數值）→ 驗證結果（量測/複檢）

來源

來源	情境數
自建情境（Vicuna 式生成）	10
IRWOZ 2.0 / IRWoZ 1.0（Aalborg Univ.）	4
KIDE4I（Tekniker Foundation, Industry 5.0）	2
ICNLSP 2023 工業技術文件語料	1
NEU-DET（熱軋鋼帶表面缺陷）	3
DeepPCB（PCB 表面缺陷）	3
MVTec AD（通用工業異常）	6
KolektorSDD（金屬表面細微裂紋）	1

Schema（JSONL，每行一筆）

{
  "id": "hmi-18",
  "domain": "鋼鐵/金屬",
  "machine": "熱軋鋼帶線",
  "scenario": "情境描述",
  "source": "NEU-DET（熱軋鋼帶表面缺陷）",
  "transcript": [
    {"role": "operator", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "model": "gemma3:27b",
  "ts": "2026-08-09 19:00:00"
}

生成方式

scripts/gen_hmi_dataset.py：seed 情境 → ollama（gemma3:27b）產生 6~12 輪 transcript，驗證角色合法、內容非空。

隱私保護

合成內容：所有情境與對話皆為 LLM 生成，非真實工廠事件或真實人員對話。
無個人資料：transcript 中不包含真實姓名、電話、Email、地址等個人識別資訊；操作員一律以「operator」角色呈現。
無機台敏感資料：對話中的參數數值為合成情境設定，不代表任何真實產線的實際參數。

授權與版權

本資料集：CC BY 4.0。可自由使用、散布、改作（含商業使用），惟須姓名標示。CC 4.0 §4 並涵蓋資料庫之 sui generis 權利。
附註：資料來源、上游條款與安全提醒見 NOTICE.md。該等內容為告知性質，不對 CC BY 4.0 附加任何條件。
情境來源：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（例如「MVTec AD metal_nut 類」）。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。致謝見 NOTICE.md。
生成模型：gemma3:27b。依 Gemma 條款 §1.1(e)/§3.3，Outputs 不受 §3.1 散布條件拘束。惟若您微調出近似 Gemma 之模型，該模型本身可能構成 Model Derivative —— 見 NOTICE.md §2。
⚠️ 非工程指引：對話參數值為合成示例。未經負責工程師獨立驗證，請勿用於驅動真實機台控制或安全連鎖。

著作人與引用

Tzer-jen Wei

引用格式同英文段落之 BibTeX。