The dataset viewer should be available soon. Please [retry later].

#

	Sinabis Green Energy – Synthetic Email Dataset

More than 10 million emails and 1.4 million calendar files from the everyday business of a fictional renewable-energy group. They span ten and a half years, six countries and 29 languages, and they include real attachments. Every person, company, authority, law and event is made up. The content was written by large language models, and the mailgen generator built the structure around it.

All figures in this file were computed from the files themselves (as of 7 Oct 2026). They are not target values from the configuration.

Also available: the smaller Sinabis Oil – Synthetic Email Dataset (1.25 million emails, four languages) from a different fictional company, an oil and gas group.

#

	At a glance

| Emails | 10,088,914 (177.8 GB,.eml per RFC 5322): 8,768,999 from run 1, 1,319,915 from run 2 | | Calendar files | 1,420,492.ics (iCalendar, RFC 5545): 492,543REQUEST , 908,716REPLY , 19,233CANCEL | | Business cases | 3,680,231 with emails: 2,013,994 email threads, 851,530 single emails, 806,833 system notifications, 7,874 HR lifecycle events | | Meetings | 322,673 meetings with 1,269,964 calendar files (.ics only, no email) | | Attachments | 2,227,176 files in 2,075,531 emails: 1,158,365 PDF · 507,520 DOCX · 450,989 XLSX · 110,302 PNG/images | | Time span | December 2015 to July 2026, growing from 0.54 to 1.34 million emails per year | | Languages | 29: 87.4 % English, 8.8 % Spanish, the rest spread over 27 languages | | People | 7,850 employees (2,880 active in 2016 → 7,310 in 2026) in 21 departments, 9,663 external contacts | | Organisations | 7 group companies + 832 others (customers, suppliers, subcontractors, authorities, NGOs, media …) |

#

	What it is about

Sinabis Green Energy (headquartered in Austin, Texas, fictional) builds and operates everything that is green: rooftop and utility-scale solar (installation and annual maintenance), wind farms, EV charging parks, autonomous shuttles, battery storage, water treatment and CO2-neutral buildings. A foundation cleans oil from the sea and recycles plastic bottles, a research institute publishes (made-up) studies, and a government affairs office lobbies made-up politicians for made-up US funding programmes.

There are 830 topics in the emails, including:

solar annual maintenance renewal · inverter fault work order · yield below forecast · residential underperformance complaint · subcontractor liability dispute · scaffolding subcontract redlines · module short shipment · battery connector recall · shuttle brake sensor recall · boil-water advisory · net-metering application delay · fall-protection near miss · heat-wave crew warning · grant applications and lobbying · HR, IT, finance and procurement topics and more.

The COVID-19 pandemic is part of the story from 2020 on (masks, remote work, supply crisis): 586,408 emails (5.8 %) relate to it. 484,073 emails (4.8 %) are private or informal exchanges between colleagues. Single emails are newsletters, memos, circulars, press releases and payment reminders; system notifications come from the service desk, HR portal, accounts payable, procurement, facilities and shipping.

	Who is writing?

Group	Count	Examples
Sinabis Green Energy group	7 companies, 16 sites, 11 business units, 21 departments	Inc. (USA), Japan K.K., Australia Pty, NZ Ltd, Shanghai Co., Middle East FZ-LLC, Ocean & Earth Foundation
Business customers and utilities	188 + 29	Cedar Mesa Municipal Utilities, Pacific Coast Utilities Board, Piedmont Power Co-op
Suppliers and subcontractors	150 + 143	Reclaimcraft Materials, Kauri Cable & Conduit, Polarbeam Glass & Insulation
Authorities and municipalities	50 + 25	Colorado Highlands Clean Power Authority, City of Tucson Community Power Office
Certifiers, insurers, consultants, law firms	9 + 10 + 16 + 15	Gridmark Testing & Certification, Seaway Marine & Property Insurance
NGOs, universities, media	35 + 23 + 25	Lone Star Agricultural University, Continental Press Wire
Politicians and their offices	50 + 100 staff	USA 24, Japan 6, Australia 6, NZ 5, UAE 5, China 4
Private persons	4,950 residents and customers, 830 supporters, 300 landowners

Activity is unevenly distributed, like in a real mailbox: of 7,828 internal sender addresses, the busiest 10 % write 56.5 % of all internal emails. People only appear during their period of employment; around their leaving date there is a farewell email, an HR note or a handover, and after that they never write again.

	Business units

Unit	Emails	Share
Solar	2,016,753	20.0 %
Corporate	1,434,396	14.2 %
Charge	1,100,095	10.9 %
Wind	1,052,936	10.4 %
Mobility	875,705	8.7 %
Buildings	694,278	6.9 %
Water	681,516	6.8 %
Storage	606,146	6.0 %
Government Affairs	580,871	5.8 %
Foundation	559,763	5.5 %
Research	486,455	4.8 %

Threads refer to 2,883 projects and 470 assets (wind farms, solar parks, charging parks, storage, water works, buildings, AV fleets) over months: same numbers, same team, same customer.

#

	Language distribution

The language of each email is recorded in the manifest and in the Content-Language header.

Language	Emails	Share
English	8,817,650	87.4 %
Spanish	887,504	8.8 %
Japanese	140,330	1.4 %
Chinese	109,678	1.1 %
Arabic	89,164	0.9 %
German	6,634	0.1 %
Korean	5,703	0.1 %
Hindi	4,569	< 0.1 %
Italian	3,173	< 0.1 %
French	3,114	< 0.1 %
Portuguese	3,110	< 0.1 %
18 more (nl, pl, id, tr, ru, tl, vi, th, he, ro, nb, da, uk, el, sv, cs, hu, fi)	18,285	0.2 %
Total	10,088,914

Every email is monolingual: subject, body, signature, disclaimer and history header are all in the same language.

Attachments by language

		DOCX	XLSX	Images
English	1,026,458	417,334	403,198	96,395
Spanish	116,470	46,390	33,368	10,293
Japanese	4,247	18,439	5,193	1,506
Chinese	3,407	15,575	3,940	1,061
Arabic	3,458	6,977	3,755	839
Others	4,325	2,805	1,535	208

#

	Models used

The model that wrote each email is recorded in manifest.jsonl (fields model and worker).

Model	Emails	Share
Gemma 4 26B-A4B (NVFP4)	6,214,888	61.6 %
Qwen3.6-35B-A3B (AWQ, 4-bit)	3,045,491	30.2 %
DeepSeek V4 Flash	15,113	0.1 %
Qwen 3.6 / Qwen3.5-9B	6,589	0.1 %
no LLM , rule-based (system notifications)	806,833	8.0 %

Across both runs the models read 4.75 billion and wrote 4.40 billion tokens. These totals also include the small share spent on the calendar emails that are not part of this folder. For rule-based emails the model field names the worker that handled the case, but no language model was involved. The .ics files are built by rules, without a model.

The PDF, DOCX and XLSX attachments are real files. Invoices, quotes, purchase orders and measurement reports are built without a model, with correct totals and the same numbers as in the email text; contracts, terms and conditions, reports and studies get their text from the model. The PNG attachments are charts, trend plots and screenshots rendered by the generator, plus a small number of photos from an image model.

#

	Downloads

The dataset is split into nine zip archives (about 101 GB zipped, about 187 GB unpacked including the manifests; the .eml files alone are 177.8 GB), so you can start with a single part and do not need to fetch everything at once.

Archive	Size	Contents
Sinabis_10M_emails_part_01.zip	15.88 GB	emails/ case folders, part 1 of 8
Sinabis_10M_emails_part_02.zip	16.15 GB	emails/ case folders, part 2 of 8
Sinabis_10M_emails_part_03.zip	12.38 GB	emails/ case folders, part 3 of 8
Sinabis_10M_emails_part_04.zip	11.39 GB	emails/ case folders, part 4 of 8
Sinabis_10M_emails_part_05.zip	11.27 GB	emails/ case folders, part 5 of 8
Sinabis_10M_emails_part_06.zip	11.23 GB	emails/ case folders, part 6 of 8
Sinabis_10M_emails_part_07.zip	11.08 GB	emails/ case folders, part 7 of 8
Sinabis_10M_emails_part_08.zip	10.67 GB	emails/ case folders, part 8 of 8
Sinabis_10M_ics.zip	971.7 MB	ics/ : 1,420,492.ics files

The eight email parts hold the emails/sNNNNN/ case folders in ascending order, each folder with its manifest.jsonl. To get the complete tree, unpack all archives into the same directory (on Windows and macOS, merge the folders when asked). Every archive also works on its own.

hf download sinabis-group/sinabis-10m-synthetic-emails \
  Sinabis_10M_emails_part_01.zip --repo-type dataset --local-dir .
unzip -q Sinabis_10M_emails_part_01.zip

hf download sinabis-group/sinabis-10m-synthetic-emails \
  --repo-type dataset --local-dir sinabis-10m
cd sinabis-10m && for z in *.zip; do unzip -q "$z"; done

#

	Folder structure

In the repository:

.
├── README.md
├── Sinabis_10M_emails_part_01.zip … part_08.zip    → emails/
└── Sinabis_10M_ics.zip                             → ics/

After unpacking all archives into one directory:

.
├── emails/                 10,088,914 .eml
│   ├── s00000/
│   │   ├── 000000000-01.eml
│   │   ├── …
│   │   └── manifest.jsonl
│   ├── …
│   └── s04040/
└── ics/                    1,420,492 .ics
    ├── s00000/
    │   ├── 000000001-01.ics
    │   └── …
    ├── …
    └── s04040/

The sNNNNN folders hold 1,000 business cases each and only split the data so that no single folder holds millions of files. They carry no meaning. An .ics file lives in the folder with the same sNNNNN name as its case.

Each emails/sNNNNN folder contains a manifest.jsonl with one line per file: file, date, sender, recipients, subject, kind (thread, single, meeting, notification, lifecycle), language, topic, business unit, site, project, asset, counterparty, attachments, COVID flag and model.

Manifest note: The manifests were not rewritten when the folder was reorganised. Besides the 10,088,914 lines for the emails here they still contain 1,236,641 lines for the calendar emails that were moved out, and 183,851 lines for the .ics files of run 2 ("format": "ics", path emails/sNNNNN/…ics, the file is now in ics/sNNNNN/). Only count lines whose file exists.

#

	How everything is linked

	Email threads

<case>-<NN>.eml
 │      └── position in the thread (01, 02, …)
 └───────── 9-digit case number, same for all emails of the case

• All emails of one case share the same file name prefix .
• The headers link the emails: Message-ID ,In-Reply-To andReferences each point to the previous email.
• Every email contains the quoted history of all earlier emails, either in Outlook style or> style depending on the sender's mail client, so a single file already shows the whole thread.
• 1,166,273 first emails (31.8 % of all first emails) also carry an HTML version (multipart/alternative ).
• 99.7 % of threads have 2 to 8 emails , 77.5 % have 3 to 6.
• Gaps: In 16,943 threads of run 1 a position is missing, because that email contained a meeting invitation and was moved out with the other calendar emails (not part of this release).

	Attachments

• Attachments are embedded in the email (MIME, Base64), not stored as separate files next to it. 2,075,531 emails (20.6 %) carry at least one.
• The document is the source the email refers to , for example an invoice, a maintenance report or a contract draft, not a copy of the email text. It has the same language and the same date as the email.
• No email in emails/ contains atext/calendar part or an.ics attachment.
• Meeting links (meet.sinabis-green.com ) and team sites (teams.sinabis-green.com ) in the text are fictional.

	Meetings and calendar files (`ics/`)

ics/sNNNNN/<case>-<NN>.ics

• A meeting consists of several .ics files with the same case number: the invitation (REQUEST ) to all attendees, acceptances or declines (REPLY ) to the organiser, and possibly an update or cancellation (CANCEL ). 322,673 meetings produce 1,269,964 files this way, 3.9 per meeting on average.
• The remaining 150,528 files are invitations that came up inside an email case (110,332 from threads, 40,196 from single emails).
• Where the files come from:

Source	Files	Matching email
Run 1, extracted from the text/calendar part of the calendar emails	1,236,641	not part of this release (the calendar emails were moved out)
Run 2, meeting invitations written directly as .ics	164,238	none, there never was an email
Run 2, invitations at the end of a thread or single email	19,613	emails/sNNNNN/<same name>.eml

• The files use CRLF line endings, METHOD ,UID ,ORGANIZER andATTENDEE , so they can be imported directly into calendar software.

#

	How timestamps are created

Timestamps are set by the generator, not by the language model.

1. Start time: For each case the generator picks a day between 1 Jan 2016 and 30 Jun 2026 and a time within office hours. The number of cases per year grows with the workforce.
2. Thread: Replies always come after the previous email, with plausible gaps between them.
3. Time zones: TheDate header is in thesender's local time zone , including daylight saving time (e.g.-0500 in Houston,+0900 in Tokyo,+1200 /+1300 in Auckland).
4. People: Every person has an employment or business-relationship window. Someone hired in 2024 does not send emails in 2019.
5. Time-bound topics: Topics such as COVID-19 (2020–21) or the supply crisis (2021–22) only appear in their period.
6. Attachments carry the date of their email.

Emails per year (local date of the Date header)

2016	2017	2018	2019	2020	2021	2022	2023	2024	2025	2026 (to July)
543,229	625,167	713,299	792,423	882,984	981,067	1,067,436	1,160,651	1,260,646	1,344,175	717,834

#

	Notes

• Fictional, and recognisably so: All names, companies, addresses, phone numbers, authorities, programmes, laws, politicians and events are made up. External domains use the reserved TLD.example , authorities*.gov.example ; phone numbers are in the fictional555-01xx range. Laws and programmes carry recognisably invented names such as "Windbreak Jobs Act (H.R. 48812)". Real cities appear as locations only.
• Fake filter: After every generation step a filter replaces real brands, authorities, laws, politicians, celebrities, universities and media, and rewrites any real domain to.example .
• Marked as synthetic: All 10,088,914.eml files carry the headerX-Synthetic-Data: fictional .
• Watchdog: Every written email was decoded and checked for foreign scripts (e.g. Chinese characters in English emails), placeholders such as[Your Name] , model artefacts and real names. Failing cases were regenerated.
• Edge dates: 3 emails are dated 31 December 2015 (local time) because of time zone shifts at the start of the time span; the latest email is from 12 July 2026.
• Manifest duplicates: 38 manifest lines are repeated entries from cases that were regenerated; the.eml file exists only once.

#

	Methodology

The prompt templates, topic catalogue, world definition and fake filter (prompts.json, topics.json, world.json, guard.json) were used to prompt the LLMs; they are not part of this release. The company world is generated deterministically from seed 7.

#

	License

This dataset is released under CC BY 4.0. You may copy, modify, distribute and use it for any purpose, including commercially, as long as you give appropriate credit, for example:

Sinabis GmbH (2026). Sinabis Green Energy – Synthetic Email Dataset. https://huggingface.co/datasets/sinabis-group/sinabis-10m-synthetic-emails

The email content was generated using Gemma 4 and Qwen 3.6 (open weights) and DeepSeek V4 Flash. Generated text is treated as free of third-party rights; the underlying models' own terms of use apply to the act of generation, not to redistribution of this dataset.

#

	Citation

If you use this dataset, please cite it as:

Sinabis GmbH (2026). Sinabis Green Energy – Synthetic Email Dataset.
https://huggingface.co/datasets/sinabis-group/sinabis-10m-synthetic-emails

BibTeX:

@misc{sinabis_green_email_dataset_2026,
  author = {{Sinabis GmbH}},
  title  = {Sinabis Green Energy -- Synthetic Email Dataset},
  year   = {2026},
  url    = {https://huggingface.co/datasets/sinabis-group/sinabis-10m-synthetic-emails}
}

• Downloads last month
• 11

source & further reading

huggingface.co — original article How Should Multiple AI Agents Work Together? Download speed way too slow Keeping the Attention Sink Exact in Activation-Whitened SVD Compression