aiDotEngineer/ai-engineer-datasets

AI Engineer Conference Talks: the applied-AI engineering corpus

Private organizer-controlled dataset. Organizer-authorized noncommercial model training and evaluation are permitted for these transcripts; public distribution, commercial use, audio release, and image redistribution each require separate organizer approval.

A first-party AI Engineer conference-talk dataset with timestamped technical transcripts and publicly grounded metadata.

The underlying conference talks are already publicly available on the official AI Engineer YouTube channel. This private organizer-controlled dataset adds high-quality provider-generated retranscription, structured public-source metadata, word- and character-level timestamps, grounded topics and entities, summaries, chapters, and recording-level joins. It does not include source audio.

Important limits: Transcripts, derived annotations, and benchmark candidates are machine-generated; annotation and evaluation share a generator and can share its biases. There is no human gold or measured accuracy. Automated privacy filtering is best-effort, timing comes from the transcription provider, and source audio is not included.

Conference talks	Recorded hours	Timestamped words	Transcript segments	Entity mentions	Conference editions
1,041	448.38	4,663,555	147,853	82,340	7

252,918 rows across 8 purpose-built configurations; counts include multiple complementary views of the same talks, not independent recordings. Median recording length: 19.05 minutes across 7 conference editions (2023–2026).

Start with the data

Access requires membership in the private AI Engineer Hugging Face organization and a signed-in Hugging Face account with permission to this repository. Install the datasets client and authenticate once:

pip install -U datasets
hf auth login

Named configurations load as native Parquet datasets. Talks, transcript segments, and retrieval examples join using stable talk and segment identifiers:

from datasets import load_dataset

repo = "aiDotEngineer/ai-engineer-datasets"
talks = load_dataset(repo, "talks", split="train", token=True)
segments = load_dataset(repo, "segments", split="train", token=True)
retrieval = load_dataset(repo, "retrieval", split="train", token=True)
benchmark = load_dataset(repo, "benchmark", split="test", token=True)

print(talks.num_rows, segments.num_rows)  # 836, 114905
print(segments.features["word_timings"])  # native nested word + character timings

talk_id = talks[0]["id"]
talk_segments = segments.filter(lambda row: row["talk_id"] == talk_id)

For this release, the training views contain 836 talks and 114,905 timestamped segments. The benchmark intentionally has no training split.

Eight joined views of the same corpus

Configuration	Total rows	Train	Validation	Test	What it contains
talks	1,041	836	94	111	Full talks, public presenters, conference metadata, and presentation dates
segments	147,853	114,905	13,102	19,846	Timestamped transcript passages with native nested word and character timing
topics	5,932	4,763	536	633	Publicly grounded technical topic labels and generation provenance
entities	82,340	65,288	7,601	9,451	Publicly grounded technical entity mentions and aliases
summaries	1,041	836	94	111	Grounded talk-level summaries linked to transcript passages
chapters	5,697	4,563	510	624	Timestamped chapter boundaries linked to transcript passages
retrieval	5,960	4,792	534	634	Synthetic queries, relevant passages, and split-safe hard negatives
benchmark	3,054	—	1,401	1,653	Held-out machine-generated retrieval, entity, and topic candidates

All derived records inherit their source talk's partition. Each of 3 reviewed repeated-presentation families stays within one partition; unknown duplicate presentations have not been exhaustively ruled out. Retrieval negatives cannot cross a split or reuse another recording from the same known presentation family. Held-out benchmark examples are machine-generated review candidates, not human-gold annotations.

Conference coverage

7 conference editions (2023–2026) contribute 1,041 speech recordings. Event assignment and session identification are separate claims: every event is assigned, while session matches are separately reported as first-party exact or machine-corroborated.

Conference edition	Recordings	Share of corpus
AI Engineer World's Fair 2025	298	28.6%
AI Engineer World's Fair 2026	270	25.9%
AI Engineer Europe 2026	187	18.0%
AI Engineer World's Fair 2024	112	10.8%
AI Engineer Code 2025	67	6.4%
AI Engineer Summit 2025	67	6.4%
AI Engineer Summit 2023	40	3.8%

How much material is here?

448.376 hours of original recording time, approximately 18.7 days, contain 4,663,555 timestamped words and 20,789,178 character-level timestamp pairs across 147,853 transcript segments. Recording time includes pauses and non-speech events; it is not a claim of speech-only duration. The median talk lasts 19.05 minutes.

Recording length	Recordings	Share of corpus
5–10 min	64	6.1%
10–20 min	546	52.4%
20–30 min	270	25.9%
30–45 min	45	4.3%
45–60 min	30	2.9%
60–90 min	51	4.9%
90+ min	35	3.4%

Grounded enrichment, with a visible paper trail

Each recording receives organizer-aware research before retranscription and a second, publicly web-grounded research pass after the improved transcript is available. Generated topics, entities, summaries, chapters, and retrieval examples point back to supporting transcript segments; unsupported candidate claims are excluded rather than promoted to facts.

Research measure	Observed value
Post-transcription, web-grounded recordings	1,041
Completed second-pass public web searches	1,363
Unsupported named-entity candidates omitted	115
Unsupported retrieval candidates omitted	16
Human review of generated research	Not completed

Generated annotations remain synthetic and unreviewed. A public-source citation or technical label supports research traceability; it does not make a model-generated assertion equivalent to human verification.

What technical problems do the talks cover?

These categories use the same governed category and subcategory assignments as the topic rows and taxonomy chart. Each topic row belongs to one category, while a talk can appear in multiple categories; distinct-talk shares always use the complete recording count.

Technical category	Topic rows	Distinct talks	Share of talks
Developer tools	1,050	515	49.5%
RAG, context, and search	876	483	46.4%
Evals	424	267	25.6%
Infrastructure and deployment	413	272	26.1%
Agent engineering	380	291	28.0%
Safety and governance	372	262	25.2%
APIs, MCP, and protocols	336	230	22.1%
Architecture	309	246	23.6%
Leadership	282	189	18.2%
Reasoning and models	263	218	20.9%
Enterprise	253	174	16.7%
Speech and audio	162	77	7.4%
Data and model adaptation	159	100	9.6%
Observability and reliability	131	107	10.3%
Creative and generative media	107	43	4.1%
Vision and video	101	80	7.7%
Industry applications	92	66	6.3%
Finance	78	50	4.8%
Healthcare	68	30	2.9%
Robotics and world models	55	24	2.3%
Other / unclassified	21	19	1.8%

Developer tools contains Coding agents and assistants (444 topic rows; 299 talks), Developer workflows, testing, and review (349 topic rows; 240 talks), IDEs, SDKs, and developer platforms (257 topic rows; 201 talks).

Grounded entity mentions

Entity type	Grounded mentions
concept	38,209
product	22,997
organization	10,316
person	10,290
repository	378
paper	150

Entity rows describe grounded occurrences spanning technical concepts, commercial and open-source products, organizations, public figures, code repositories, and research papers. Person mentions are not counts of unique people; organization mentions are not counts of unique companies.

Publicly listed speakers and companies

Official conference rosters and public-source research identify 1,030 distinct named presenters and 610 distinct presenter-affiliated organizations across 1,031 recordings, including 987 recordings with a listed affiliation. This affiliation-only organization count can include universities and research institutes and reflects historical conference context, not inferred current employment.

Across presenter affiliations and companies that are a clear central subject of the public talk title, official playlist, or track, the corpus represents 691 distinct companies across 976 talks. Each canonical company is counted at most once per talk; incidental transcript mentions, generic social platforms, and universities are excluded.

Canonical speaker headshots

761 distinct event-scoped organizer speaker identities have a verified official conference portrait, covering 811 presenter appearances across 696 recordings.

The talks rows expose presenters[].canonical_photo_url, presenters[].alternate_photo_urls, and aligned presenters[].alternate_photo_metadata together with exact photo_source_event_id, photo_source_person_id, photo_source, photo_rights_status, source-checksum, review, and opt-out metadata. Canonical portraits use event-qualified organizer speaker identifiers and first-party conference speaker records; unsupported name matches, attendee guesses, and inferred employment are excluded.

The full first-party source audit covered 1,607 organizer identities and 1,469 restricted portrait sources across 8 conference rosters.

The same organizer records link 972 verified public professional-profile URLs. Emcees, moderators, hosts, and speakers retain their organizer-confirmed participation roles when available.

The release additionally includes 3,422 distinct verified conference gallery photos attributed to 161 exact organizer-identified speakers, with 3,952 gallery-photo links across 145 talks. Gallery references are limited to speaker-attributed first-party conference photographs; unknown attendees and time-only guesses are excluded.

The complete reviewed conference gallery contains 7,213 preserved person–photo assignments across 607 known identities. Its 4 exact-photo-only exceptions remain displayable but cannot seed identity matching; no verified gallery assignments were discarded.

Conference	Named presenters	Verified portraits	Conference gallery photos
AI Engineer World's Fair 2026	320	229	3,422
AI Engineer World's Fair 2025	350	151	0
AI Engineer Europe 2026	198	127	0
AI Engineer World's Fair 2024	141	127	0
AI Engineer Code 2025	76	49	0
AI Engineer Summit 2025	81	41	0
AI Engineer Summit 2023	46	37	0

Headshots and verified conference gallery photos are HTTPS references, not bundled image files. Every photo is marked restricted; linking to a photo does not grant a license or authorize image redistribution, commercial use, or model training. No image bytes or biometric descriptors are included.

Public-source metadata coverage

Measure	Recordings	Share of talks
First-party recording-to-session matches	207	19.9%
Publicly corroborated recording-to-session matches	359	34.5%
Recordings without an exact or corroborated session match	475	45.6%
Observed presentation dates	511	49.1%
Organizer-inferred online-track presentation dates	215	20.7%
Unavailable presentation dates	315	30.3%
Observed YouTube upload dates (separate from presentation dates)	1,041	100.0%
Observed YouTube publication timestamps	1,041	100.0%
Recordings with publicly identified presenters	1,031	99.0%
Recordings with publicly identified presenter affiliations	987	94.8%
Recordings with official organizer-hosted presenter headshots	696	66.9%

An exact organizer recording ID, a machine-corroborated schedule match, an observed presentation date, an organizer-inferred online-track date, and a YouTube upload timestamp are different kinds of source information. Online-track dates remain explicitly inferred; upload dates never substitute for presentation dates.

Research, editorial, and model-development value

• Editorial discovery: locate relevant public talks and timestamped passages across seven conference editions.
• Conference research: study grounded topic labels, public speakers, organizations, products, and session metadata.
• Archive navigation: connect full transcripts, chapters, summaries, and detailed word/character timings without redistributing audio.
• Provenance review: distinguish official presenter rosters, observed session matches, inferred dates, and synthetic annotations.
• Model development and evaluation: use the organizer-authorized private transcript corpus for noncommercial training, retrieval, and held-out testing.

Held-out evaluation candidates

The benchmark contains 3,054 machine-generated examples: 1,401 validation and 1,653 test. There is no benchmark training split, and no case from a training talk is included.

Evaluation task	Held-out candidates
Transcript retrieval	1,018
Technical entity extraction	1,025
Technical topic classification	1,011

The generator excluded 74,843 training-source rows before constructing held-out candidates and capped each task at 5 examples per talk.

These are synthetic review candidates, not human-gold labels or adjudicated relevance judgments. Human-reviewed benchmark rows: 0. Word error rate (WER), diarization error rate (DER), Jaccard error rate (JER), other transcription or diarization accuracy, and benchmark scores have not been measured.

Separate transcription-system comparison

Our AI Engineer transcription and diarization benchmark is a separate historical comparison of five speech-to-text systems across 100 talks. Its silver-reference disagreement scores were Speechmatics 8.98, NVIDIA Parakeet v3 9.19, AssemblyAI 10.19, NVIDIA Parakeet v2 10.99, and Deepgram 12.42; lower disagreement is better against that particular silver reference.

The comparison has no human-adjudicated gold transcript and does not establish this dataset's word error rate or diarization accuracy.

Transcription provenance and limits

All speech recordings are English and received high-quality machine retranscription. Word and character timings retain observed absolute millisecond positions as native nested Parquet arrays. Public presenter metadata does not establish who spoke each individual passage. Human-measured transcription and diarization accuracy are unavailable.

Provider-observed non-speech audio-event timestamps use a controlled label set; 8,954 events were observed across 924 recordings. Event labels are metadata only; original audio is never included.

The original inventory contains one instrumental/non-speech recording that is preserved privately and excluded from this speech-only release. Source audio, provider responses, normalized source transcripts, sensitive-entity quarantine, operational receipts, credentials, and JSONL intermediates are never included in the dataset.

Privacy-preserving release

No unresolved sensitive transcript leakage was reported by the release validator.

Privacy measure	Count
Sensitive entities masked	822
Talks containing a masking operation	311
Segments containing a masking operation	2,261
Masked words	4,781
Masked characters	24,978
Unresolved sensitive-text validation issues	0

Masking applies to exported views only. Original source artifacts remain immutable, and restricted provider quarantine stays outside the Hugging Face bundle.

Charts describe the completed speech-only release and contain aggregate counts only; they never expose transcript text, private identifiers, named speaker lists, or audio.

Access and permitted use

This is a private, organizer-controlled dataset. Organizer-authorized noncommercial model training and evaluation of the transcript corpus are permitted. Public redistribution, commercial use, source-audio release, image redistribution, biometric identification, and any open-license declaration require separate organizer approval and are not granted here.

Transcript authorization does not sublicense third-party websites, external materials, or separately hosted conference photographs; their independent restrictions remain in force.

Corrected transcript layer

The separate corrected_transcripts configuration contains 1,041 recordings and 147,853 source-linked segments. It carries 11,244 lexical edits and 1,697 separate speaker annotations; 2,847 candidates remain unresolved. Model-adjudicated corrections are unreviewed, not human gold. Original and corrected text retain privacy masking; source voice-cluster IDs are unchanged. Speaker annotations are separate text-context inferences, not verified voice identities. Timings inherit the original video clock, not new forced alignment.

9 recordings have 11 extra observed event labels in the original full-text representation relative to the word stream. The original view preserves those labels; the corrected view follows the verified word stream. These representation differences are not counted as lexical edits.

The eight pre-existing configurations retain their original source content, with private generation labels removed from exported identifiers: entities, summaries, chapters, retrieval and benchmark rows have not been regenerated from corrected text. Their charts and aggregate statistics describe that frozen snapshot. The new layer does not change rights or grant public distribution, commercial use, audio, or image permissions.

Exported identifiers use opaque deterministic IDs where necessary to withhold private generation labels. No talk content, transcript text, timestamps, row counts, or entity labels were changed by this publication cleanup.