Datasets:
release_id stringdate 2026-11-20 00:00:00 2026-11-20 00:00:00 | doc_id stringlengths 13 13 | title stringlengths 2 378 ⌀ | abstract stringclasses 0
values | authors listlengths 1 504 ⌀ | published_date date32 | published_precision stringclasses 3
values | doc_type stringclasses 11
values | is_withdrawn bool 1
class | canonical_url stringlengths 25 162 | field_sources listlengths 0 7 | identity dict | content dict | release dict | annotations dict |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2026-11-20 | s001jbmrg8nn6 | Malicious Image Analysis via Vision-Language Segmentation Fusion: Detection, Element, and Location in One-shot | null | [
{
"name": "Hang Sheng",
"orcid": "0000-0002-0597-618X",
"affiliation_raw": "Shanghai Jiao Tong University"
},
{
"name": "Chaoxiang He",
"orcid": "0009-0008-8936-9336",
"affiliation_raw": "Shanghai Jiao Tong University"
},
{
"name": "Hongsheng Hu",
"orcid": null,
"affiliat... | 2025-12-04 | day | preprint | false | https://arxiv.org/abs/2512.04599 | [
{
"field": "title",
"source_id": "s001jbmrg8nn6",
"origin": "provider"
},
{
"field": "abstract",
"source_id": "s001jbmrg8nn6",
"origin": "provider"
},
{
"field": "authors",
"source_id": "s001jbmrg8nn6",
"origin": "provider"
},
{
"field": "authors",
"source_id"... | {
"anchor_source_id": "s001jbmrg8nn6",
"channels": [
"openalex",
"semantic_scholar"
],
"former_doc_ids": [],
"override_ids": [],
"members": [
{
"source_id": "s001jbmrg8nn6",
"channel": "openalex",
"channel_class": "open",
"key_kind": "doi",
"key_value": "10.48550/ar... | {
"source_id": "s001jbmrg8nn6",
"role": "html",
"quality": "flagged",
"source": {
"url": "https://arxiv.org/html/2512.04599v1",
"sha256": "61ad2168bc2f47af1f3f3c60c3a7c401b4e50cfc52acd75728b04de8c6ea473a",
"bytes": 287366,
"media_type": "text/html",
"captured_at": "2026-10-22T08:39:40.676000... | {
"abstract": "withheld_pending_rights_review",
"text": "not_distributed",
"quote_chars": 2648,
"cuts": []
} | {
"screen": {
"metadata": {
"kind": "screen",
"version": 1,
"status": "ok",
"error_class": null,
"config_id": "screen:screen-v12:gpt-6-luna:63dc7d28",
"decided_at": "2026-10-22T13:42:27.238000",
"origin": "ai_check",
"annotator": "gpt-6-luna",
"provenance": "p... |
2026-11-20 | s001npd5252t6 | Unbounded Embedded Agency: AEDT w.r.t. rOSI | null | [
{
"name": "Cole Wyeth",
"orcid": null,
"affiliation_raw": null
}
] | 2025-07-20 | day | forum_post | false | https://www.lesswrong.com/posts/B6gumHyuxzR5yn5tH | [
{
"field": "title",
"source_id": "shpn035bz8qtc",
"origin": "provider"
},
{
"field": "authors",
"source_id": "shpn035bz8qtc",
"origin": "provider"
},
{
"field": "published_at",
"source_id": "shpn035bz8qtc",
"origin": "provider"
},
{
"field": "doc_type",
"sourc... | {
"anchor_source_id": "s001npd5252t6",
"channels": [
"alignmentforum",
"lesswrong"
],
"former_doc_ids": [],
"override_ids": [],
"members": [
{
"source_id": "s001npd5252t6",
"channel": "lesswrong",
"channel_class": "open",
"key_kind": "fm_post",
"key_value": "B6gumHy... | {
"source_id": "swmq9aqetff6c",
"role": "post",
"quality": "ok",
"source": {
"url": null,
"sha256": "dc1655d3ece1c06b7741cece5739128976d040a4fdac13d8c0bddf43c25ccb45",
"bytes": 469665,
"media_type": "text/html",
"captured_at": "2026-10-17T09:36:59.017000",
"revision": null
},
"text":... | {
"abstract": "none",
"text": "not_distributed",
"quote_chars": 2290,
"cuts": []
} | {
"screen": {
"metadata": {
"kind": "screen",
"version": 1,
"status": "ok",
"error_class": null,
"config_id": "screen:screen-v12:gpt-6-luna:63dc7d28",
"decided_at": "2026-10-17T17:55:22.152000",
"origin": "ai_check",
"annotator": "gpt-6-luna",
"provenance": "p... |
2026-11-20 | s002jdfjm1k5j | ColorGrid: A Multi-Agent Non-Stationary Environment for Goal Inference and Assistance | null | [
{
"name": "Andrey Risukhin",
"orcid": null,
"affiliation_raw": "University of Washington"
},
{
"name": "Kavel Rao",
"orcid": null,
"affiliation_raw": "University of Washington"
},
{
"name": "Ben Caffee",
"orcid": null,
"affiliation_raw": "University of Washington"
},
... | 2025-01-17 | day | preprint | false | https://arxiv.org/abs/2501.10593 | [
{
"field": "title",
"source_id": "s002jdfjm1k5j",
"origin": "provider"
},
{
"field": "abstract",
"source_id": "s002jdfjm1k5j",
"origin": "provider"
},
{
"field": "authors",
"source_id": "s002jdfjm1k5j",
"origin": "provider"
},
{
"field": "published_at",
"sourc... | {
"anchor_source_id": "s002jdfjm1k5j",
"channels": [
"openalex"
],
"former_doc_ids": [],
"override_ids": [],
"members": [
{
"source_id": "s002jdfjm1k5j",
"channel": "openalex",
"channel_class": "open",
"key_kind": "doi",
"key_value": "10.48550/arxiv.2501.10593",
"... | {
"source_id": "s002jdfjm1k5j",
"role": "html",
"quality": "ok",
"source": {
"url": "https://arxiv.org/html/2501.10593v1",
"sha256": "b2977cbc13b60f37aead22bcf3add85d4f659cb563e3ebaaa8c4d55679c39da8",
"bytes": 142603,
"media_type": "text/html",
"captured_at": "2026-10-18T01:15:43.281000",
... | {
"abstract": "withheld_pending_rights_review",
"text": "not_distributed",
"quote_chars": 1005,
"cuts": []
} | {
"screen": {
"metadata": {
"kind": "screen",
"version": 1,
"status": "ok",
"error_class": null,
"config_id": "screen:screen-v12:gpt-6-luna:63dc7d28",
"decided_at": "2026-10-18T05:58:15.554000",
"origin": "ai_check",
"annotator": "gpt-6-luna",
"provenance": "p... |
2026-11-20 | s002jkpn3ja5v | Conversational diagnostic artificial intelligence in ambulatory primary care: a prospective feasibility study | null | [
{
"name": "Peter G. Brodeur",
"orcid": "0000-0002-7391-719X",
"affiliation_raw": "Beth Israel Deaconess Medical Center"
},
{
"name": "Jacob M. Koshy",
"orcid": "0009-0005-3721-5608",
"affiliation_raw": "Beth Israel Deaconess Medical Center"
},
{
"name": "Anil K. Palepu",
"orc... | 2026-10-01 | day | paper | false | https://doi.org/10.1016/s0140-6736(26)01535-7 | [
{
"field": "title",
"source_id": "s002jkpn3ja5v",
"origin": "provider"
},
{
"field": "abstract",
"source_id": "sred7bw6ar85p",
"origin": "provider"
},
{
"field": "authors",
"source_id": "s002jkpn3ja5v",
"origin": "provider"
},
{
"field": "published_at",
"sourc... | {
"anchor_source_id": "s002jkpn3ja5v",
"channels": [
"openalex",
"semantic_scholar"
],
"former_doc_ids": [
"sred7bw6ar85p"
],
"override_ids": [],
"members": [
{
"source_id": "s002jkpn3ja5v",
"channel": "openalex",
"channel_class": "open",
"key_kind": "doi",
"k... | {
"source_id": "sred7bw6ar85p",
"role": "html",
"quality": "flagged",
"source": {
"url": "https://arxiv.org/html/2603.08448v3",
"sha256": "e87e59a728751c146310abae7bb8fe96bdcaad3f6e9275a0c279146e4d93a8be",
"bytes": 304480,
"media_type": "text/html",
"captured_at": "2026-10-16T08:06:35.541000... | {
"abstract": "withheld_pending_rights_review",
"text": "not_distributed",
"quote_chars": 2402,
"cuts": []
} | {
"screen": {
"metadata": {
"kind": "screen",
"version": 1,
"status": "ok",
"error_class": null,
"config_id": "screen:screen-v12:gpt-6-luna:63dc7d28",
"decided_at": "2026-10-16T13:21:47.634000",
"origin": "ai_check",
"annotator": "gpt-6-luna",
"provenance": "p... |
2026-11-20 | s003mwcpn5m4e | How Vision Transformers Represent Position: A Mechanistic Comparison of Absolute and Rotary Positional Encodings | null | [
{
"name": "Mahmoud Mannes",
"orcid": null,
"affiliation_raw": null
},
{
"name": "Nikhil Maturi",
"orcid": null,
"affiliation_raw": null
},
{
"name": "Parva Mehta",
"orcid": null,
"affiliation_raw": null
},
{
"name": "Jiwon Jeong",
"orcid": "https://orcid.org/0... | 2026-09-26 | day | preprint | false | https://arxiv.org/abs/2610.13992 | [
{
"field": "title",
"source_id": "s0sadsnnwwfn8",
"origin": "provider"
},
{
"field": "abstract",
"source_id": "s0sadsnnwwfn8",
"origin": "provider"
},
{
"field": "authors",
"source_id": "s0sadsnnwwfn8",
"origin": "provider"
},
{
"field": "published_at",
"sourc... | {
"anchor_source_id": "s003mwcpn5m4e",
"channels": [
"openalex",
"semantic_scholar"
],
"former_doc_ids": [],
"override_ids": [],
"members": [
{
"source_id": "s003mwcpn5m4e",
"channel": "semantic_scholar",
"channel_class": "restricted",
"key_kind": "doi",
"key_value"... | {
"source_id": "s0sadsnnwwfn8",
"role": "html",
"quality": "ok",
"source": {
"url": "https://arxiv.org/html/2610.13992",
"sha256": "525e8d9dc1605e4639b0aab5084eaa6f0cbead1fb1e0f82301e68d668cb7723a",
"bytes": 417218,
"media_type": "text/html",
"captured_at": "2026-10-22T03:36:27.856000",
... | {
"abstract": "withheld_pending_rights_review",
"text": "not_distributed",
"quote_chars": 2291,
"cuts": []
} | {
"screen": {
"metadata": {
"kind": "screen",
"version": 1,
"status": "ok",
"error_class": null,
"config_id": "screen:screen-v12:gpt-6-luna:63dc7d28",
"decided_at": "2026-10-22T13:42:27.238000",
"origin": "ai_check",
"annotator": "gpt-6-luna",
"provenance": "p... |
2026-11-20 | s003rexe6dpr4 | A Black Box for Agentic Processes: Blockchain-Anchored Evidence for AI Agent Communication, Human Oversight, and GRC Audits | null | [
{
"name": "Arslan Brömme",
"orcid": null,
"affiliation_raw": null
}
] | 2026-09-03 | day | preprint | false | https://arxiv.org/abs/2609.04017 | [
{
"field": "title",
"source_id": "s003rexe6dpr4",
"origin": "provider"
},
{
"field": "abstract",
"source_id": "s003rexe6dpr4",
"origin": "provider"
},
{
"field": "authors",
"source_id": "s003rexe6dpr4",
"origin": "provider"
},
{
"field": "published_at",
"sourc... | {
"anchor_source_id": "s003rexe6dpr4",
"channels": [
"openalex"
],
"former_doc_ids": [],
"override_ids": [],
"members": [
{
"source_id": "s003rexe6dpr4",
"channel": "openalex",
"channel_class": "open",
"key_kind": "doi",
"key_value": "10.48550/arxiv.2609.04017",
"... | {
"source_id": "s003rexe6dpr4",
"role": "html",
"quality": "ok",
"source": {
"url": "https://arxiv.org/html/2609.04017v1",
"sha256": "a4ed04c504c4f0f7487ffb99e2c6d9cca23acc8d0ba8bbcd8cea43bc03841ea0",
"bytes": 78087,
"media_type": "text/html",
"captured_at": "2026-10-15T12:31:24.956000",
... | {
"abstract": "withheld_pending_rights_review",
"text": "not_distributed",
"quote_chars": 2605,
"cuts": []
} | {
"screen": {
"metadata": {
"kind": "screen",
"version": 1,
"status": "ok",
"error_class": null,
"config_id": "screen:screen-v12:gpt-6-luna:63dc7d28",
"decided_at": "2026-10-15T19:19:26.937000",
"origin": "ai_check",
"annotator": "gpt-6-luna",
"provenance": "p... |
2026-11-20 | s004a6kwkcq8d | Vibe Coding vs. Agentic Coding: Fundamentals and Practical Implications of Agentic AI | null | [
{
"name": "Ranjan Sapkota",
"orcid": "0000-0002-5417-6744",
"affiliation_raw": "Cornell University; New York State College of Agriculture & Life Sciences"
},
{
"name": "Konstantinos I. Roumeliotis",
"orcid": "0000-0002-8098-1616",
"affiliation_raw": "University of Peloponnese"
},
{
... | 2025-05-26 | day | preprint | false | https://arxiv.org/abs/2505.19443 | [
{
"field": "title",
"source_id": "s004a6kwkcq8d",
"origin": "provider"
},
{
"field": "abstract",
"source_id": "s004a6kwkcq8d",
"origin": "provider"
},
{
"field": "authors",
"source_id": "s004a6kwkcq8d",
"origin": "provider"
},
{
"field": "published_at",
"sourc... | {
"anchor_source_id": "s004a6kwkcq8d",
"channels": [
"openalex"
],
"former_doc_ids": [],
"override_ids": [],
"members": [
{
"source_id": "s004a6kwkcq8d",
"channel": "openalex",
"channel_class": "open",
"key_kind": "doi",
"key_value": "10.48550/arxiv.2505.19443",
"... | {
"source_id": "s004a6kwkcq8d",
"role": "html",
"quality": "ok",
"source": {
"url": "https://arxiv.org/html/2505.19443v1",
"sha256": "de33aa474a6439e175ecdeb8b8913e368f4cb4ef6eb28fb9c140a2211e0a66df",
"bytes": 523694,
"media_type": "text/html",
"captured_at": "2026-10-17T15:55:48.752000",
... | {
"abstract": "withheld_pending_rights_review",
"text": "not_distributed",
"quote_chars": 2087,
"cuts": []
} | {
"screen": {
"metadata": {
"kind": "screen",
"version": 1,
"status": "ok",
"error_class": null,
"config_id": "screen:screen-v12:gpt-6-luna:63dc7d28",
"decided_at": "2026-10-17T17:55:22.152000",
"origin": "ai_check",
"annotator": "gpt-6-luna",
"provenance": "p... |
2026-11-20 | s007ezr9e45et | Zero-Shot Coordination among LLM Agents | null | [
{
"name": "Adrian Hayler",
"orcid": null,
"affiliation_raw": "University of Oxford"
},
{
"name": "Shashank Reddy Chirra",
"orcid": null,
"affiliation_raw": "University of Oxford"
},
{
"name": "Andrei Lupu",
"orcid": null,
"affiliation_raw": "University of Oxford"
},
{... | 2026-11-04 | day | preprint | false | https://arxiv.org/abs/2611.06409 | [
{
"field": "title",
"source_id": "s007ezr9e45et",
"origin": "provider"
},
{
"field": "abstract",
"source_id": "s007ezr9e45et",
"origin": "provider"
},
{
"field": "authors",
"source_id": "s007ezr9e45et",
"origin": "provider"
},
{
"field": "published_at",
"sourc... | {
"anchor_source_id": "s007ezr9e45et",
"channels": [
"openalex"
],
"former_doc_ids": [],
"override_ids": [],
"members": [
{
"source_id": "s007ezr9e45et",
"channel": "openalex",
"channel_class": "open",
"key_kind": "doi",
"key_value": "10.48550/arxiv.2611.06409",
"... | {
"source_id": "s007ezr9e45et",
"role": "html",
"quality": "ok",
"source": {
"url": "https://arxiv.org/html/2611.06409v1",
"sha256": "d699e62250eeb28bcf2d100135a4347c37c7f9a13bd896cd7ff9b70a72882a7e",
"bytes": 1186130,
"media_type": "text/html",
"captured_at": "2026-11-08T01:23:08.807000",
... | {
"abstract": "withheld_pending_rights_review",
"text": "not_distributed",
"quote_chars": 2858,
"cuts": [
{
"path": "annotations.safety_techniques.payload.techniques[0].quotes[0].text",
"chars": 175
},
{
"path": "annotations.safety_techniques.payload.techniques[1].quotes[0].text",
... | {
"screen": {
"metadata": {
"kind": "screen",
"version": 1,
"status": "ok",
"error_class": null,
"config_id": "screen:screen-v12:gpt-6-luna:63dc7d28",
"decided_at": "2026-11-08T07:16:11.117000",
"origin": "ai_check",
"annotator": "gpt-6-luna",
"provenance": "p... |
2026-11-20 | s007fg30dkjdp | Are AI Coders Snitches? An Empirical Study of Pretraining Data Detection on Code Large Language Models | null | [
{
"name": "Tianlin Li",
"orcid": "0009-0000-0307-2649",
"affiliation_raw": "Beihang University"
},
{
"name": "Yunxiang Wei",
"orcid": null,
"affiliation_raw": "Zhejiang University"
},
{
"name": "Zhiming Li",
"orcid": "0000-0002-8248-7953",
"affiliation_raw": "Peng Cheng L... | 2025-07-23 | day | preprint | false | https://arxiv.org/abs/2507.17389 | [
{
"field": "title",
"source_id": "s007fg30dkjdp",
"origin": "provider"
},
{
"field": "abstract",
"source_id": "s007fg30dkjdp",
"origin": "provider"
},
{
"field": "authors",
"source_id": "s007fg30dkjdp",
"origin": "provider"
},
{
"field": "published_at",
"sourc... | {
"anchor_source_id": "s007fg30dkjdp",
"channels": [
"openalex"
],
"former_doc_ids": [],
"override_ids": [],
"members": [
{
"source_id": "s007fg30dkjdp",
"channel": "openalex",
"channel_class": "open",
"key_kind": "doi",
"key_value": "10.48550/arxiv.2507.17389",
"... | {
"source_id": "s007fg30dkjdp",
"role": "html",
"quality": "ok",
"source": {
"url": "https://arxiv.org/html/2507.17389v2",
"sha256": "083504411c80fb2bbd532186d95a48fab0d57a38369d611ee7b5522555f9e8f3",
"bytes": 620790,
"media_type": "text/html",
"captured_at": "2026-11-01T13:55:19.954000",
... | {
"abstract": "withheld_pending_rights_review",
"text": "not_distributed",
"quote_chars": 2867,
"cuts": []
} | {
"screen": {
"metadata": {
"kind": "screen",
"version": 1,
"status": "ok",
"error_class": null,
"config_id": "screen:screen-v12:gpt-6-luna:63dc7d28",
"decided_at": "2026-11-01T15:50:38.994000",
"origin": "ai_check",
"annotator": "gpt-6-luna",
"provenance": "p... |
2026-11-20 | s007y4947by3k | IO Factory: Simulating AI-Enabled Influence Campaigns at Scale | null | [{"name":"Łukasz Olejnik","orcid":"0009-0002-3208-4707","affiliation_raw":"King's College London"},(...TRUNCATED) | 2026-08-11 | day | preprint | false | https://arxiv.org/abs/2608.10920 | [{"field":"title","source_id":"s007y4947by3k","origin":"provider"},{"field":"abstract","source_id":"(...TRUNCATED) | {"anchor_source_id":"s007y4947by3k","channels":["openalex"],"former_doc_ids":[],"override_ids":[],"m(...TRUNCATED) | {"source_id":"s007y4947by3k","role":"html","quality":"ok","source":{"url":"https://arxiv.org/html/26(...TRUNCATED) | {
"abstract": "withheld_pending_rights_review",
"text": "not_distributed",
"quote_chars": 2040,
"cuts": []
} | {"screen":{"metadata":{"kind":"screen","version":1,"status":"ok","error_class":null,"config_id":"scr(...TRUNCATED) |
The Safety Stack dataset, release 2026-11-20
The annotated document index of The Safety Stack project, with taxonomies of safety techniques and AI models in the data repository.
The Safety Stack project. Frontier labs' current alignment strategy is primarily defense-in-depth — a stack of safety techniques at every layer from data collection to deployment. The number of techniques, papers, AI models and expert opinions is vast and fast growing. Much of this information is unknown or not public, and there is no "negative map" of what is unknown.
Our project's goal is to map the space of contemporary AI safety techniques by collecting available information about the techniques, their efficiency and deployment. The goal is not just an extensive collection of relevant literature but creating a comprehensive map and showing the gaps: techniques that have not been developed, have not been deployed, are not sufficiently effective in preventing harms, or have already demonstrably failed.
This release in numbers.
- 43,558 public documents on technical AI safety: 19,253 preprints, 12,809 forum posts, 9,300 papers, 1,818 lab blog posts, 122 system and model cards, and 256 others.
- Published 2004-2026; 27,112 of them from 2025 on.
- Collected from 43 source channels: scholarly indexes (OpenAlex, arXiv, Semantic Scholar), LessWrong and the Alignment Forum, the blogs of frontier labs and AI-safety organisations, model and system card pages, a few newsletters and sources added by hand. The paper searches cover 2015 to the present.
- Selection: 153,173 records marked not relevant by triage from title and metadata (their text was not fetched); of the 80,846 documents whose text was read by the relevance screen, 43,558 were judged relevant and shipped, 37,288 were not.
- Annotated against 234 safety techniques (plus 9 group nodes for generic mentions of a field), 1,877 AI models, model families and groups, and 4,841 organisations.
We are providing an early release to make the annotated document index available so researchers can use the data and report errors while the annotations and taxonomies are being reviewed and deployment status is being investigated. The dataset builds on the Shallow Review of Technical AI Safety, a yearly map of technical AI safety research by research agenda, and is meant to supply data for its 2026 edition.
What is included. A document is included when a language model screening its text judged it relevant to at least
one of two questions: does it carry something about a concrete technique for making general-purpose AI systems safer
(the Stack question, annotations.stack_relevance), or is it technical AI safety research in the Shallow Review's
sense (annotations.shallow_review_relevance); documents borderline overall (relevant to neither question) are left
out. 30,748 documents are relevant to both questions, 5,965 only to the Stack question and 6,844
only to the Shallow Review question (the other answer irrelevant or borderline), and 1 to neither (FORMAT.md, known inconsistencies); filter on the question
you care about. Apart from sources taken by rule (posts of frontier labs, AI-safety organisations and newsletters,
sources added by hand or linked from them), a triage model judged each record from its title and metadata to decide
which full texts to fetch; the screen then read the texts. The full criteria are in the data repository's
README.
Notes on the coverage. As of this release, it is not a complete catalogue. It does not, for instance, include per-technique summaries from our taxonomy of techniques or a catalogue of evidence of use and deployment by frontier labs. There are likely relevant documents missing and some irrelevant documents included: the searches, triage and relevance screen each miss relevant documents, and the screen lets some irrelevant ones in. Counts by year or by technique describe this corpus, not the field. The 590 documents from before 2015, when the paper searches start, are 428 forum posts (LessWrong, Alignment Forum), 116 posts from organisations' blog archives, 42 works added by hand or linked from hand-added pages, and 4 others.
Annotation and metadata. The processing of the dataset has been model-driven (gpt-6-luna). The annotation of
the data records the model's reading of a document: which safety techniques from our taxonomy it discusses and in
what role (uses, evaluates, proposes, critiques, reports a lab using it); which AI models from our catalogue it names
and in what role; its document class; and, for papers with a full text, its authors' affiliations, each linked to the organisations of our organisation vocabulary it names (93.5 % of 139,114 affiliations name at least one).
These readings record what a document says, not whether it is true. The taxonomies and per-field provenance (the
producing call of every annotation, and the overrides) ship with the data; their versions and changelogs are in the
data repository.
These readings have not been individually verified, but have gone through an
Annotation quality pipeline; the section below gives their agreement with three language-model
judges. Of the 624 override rows applied to this release, 253 add a source by hand, 171 merge two records of one work, 97 fix a bibliographic field (of 96 documents), 81 set screen or relevance fields (of 27 documents), 18 are Claude stand-in annotations (of 15 documents) and 4 replace an annotation. Most of the maintainers' rows come from lists an AI assistant prepared and the maintainers approved as a whole, not row by row; each row's method says how it was prepared. Including a document in the dataset does not mean the maintainers endorse it, and
nothing in the dataset is a judgement about a lab or a technique.
Documents whose annotation came from a Claude model because the main model refused them are marked in the data
(annotator, provenance = override; 14 documents in this release). They are not in the judge samples
of the Annotation quality section.
Quick start. The dataset is one parquet file, data/corpus.parquet: one row per document, with every annotation
nested in it. DuckDB and polars read it straight from the hub. Reading a private repository needs a Hugging Face
access token with read access to it in the environment variable HF_TOKEN (and, in DuckDB, the commented line below).
DuckDB (SQL; unnest turns a document's list of techniques into one row per technique):
import duckdb
con = duckdb.connect()
# con.execute("CREATE SECRET hf (TYPE huggingface, PROVIDER credential_chain)") # private repositories: reads HF_TOKEN
con.execute("CREATE VIEW d AS SELECT * FROM 'hf://datasets/thesafetystack/the-safety-stack/data/corpus.parquet'")
# documents whose own work uses RLHF (role strength 2 or more of 4: at least "minor but real")
con.sql("""
SELECT doc_id, title, published_date, t.roles.used AS strength
FROM d, unnest(annotations.safety_techniques.payload.techniques) AS u(t)
WHERE t.slug = 'rlhf' AND t.roles.used >= 2
ORDER BY published_date DESC
""").show()
polars:
import polars as pl
d = pl.scan_parquet("hf://datasets/thesafetystack/the-safety-stack/data/corpus.parquet")
techniques = (d.select("doc_id", "title", t=pl.col("annotations").struct.field("safety_techniques")
.struct.field("payload").struct.field("techniques"))
.explode("t").unnest("t")) # one row per (document, technique)
rlhf = techniques.filter((pl.col("slug") == "rlhf") & (pl.col("roles").struct.field("used") >= 2)).collect()
Hugging Face datasets (nested fields arrive as Python dicts and lists):
from datasets import load_dataset
docs = load_dataset("thesafetystack/the-safety-stack", split="train")
def used(row, slug): # the document's own work uses the technique (strength 2 or more of 4: "minor but real")
st = row["annotations"]["safety_techniques"]
return any(t["slug"] == slug and t["roles"]["used"] >= 2 for t in (st and st["payload"]["techniques"]) or [])
rlhf = docs.filter(lambda r: used(r, "rlhf"))
Each query on the hf:// URL reads the remote file again; for repeated work download it once, with
hf download thesafetystack/the-safety-stack --repo-type dataset --revision release-2026-11-20 --local-dir stack (the whole
release, 19 files) and read stack/data/corpus.parquet. To pin this release when reading from the hub, use its tag:
hf://datasets/thesafetystack/the-safety-stack@release-2026-11-20/data/corpus.parquet, or revision="release-2026-11-20" in
load_dataset. The snippets were tested with DuckDB 2.0, polars 2.1 and datasets 5.1 (pip install duckdb polars datasets; the hf command comes with huggingface_hub).
Model mentions. annotations.model_roles.payload.models[] lists every catalogue model the name matcher found,
including the candidates the annotation rejected, which have every role 0: count a model only where
roles.mentioned >= 1 (FORMAT.md has the recipe).
Checking a download. The release ships its checker: uv run stack/tools/release-validator.py stack --rows all
checks the format version (1.0.0), every file against release.json and against its schema (every
row), the ids and the taxonomy pins; exit code 0 = valid. FORMAT.md's "Format versions" says what a format version
guarantees.
Every field is documented in FORMAT.md, with more DuckDB, polars and pandas recipes, and every value word in VOCABULARY.md.
Project, licence, corrections.
- Site: https://thesafetystack.ai; project page.
- Taxonomies, their versions and changelogs, and release formats: the
data repository. The overrides applied to this release are in
config/overrides.jsonl. - The public pipeline builds the dataset and includes its prompts and the judge evaluation.
- For errors, missing works or takedowns, use the GitHub issue forms or write privately to info@thesafetystack.ai. We acknowledge every request.
What we made — the annotations, the identity and provenance structure, the release metadata, the taxonomies and the
documentation — is released under CC0 1.0, to the extent we hold copyright or database rights in it. Third-party
material, wherever it occurs in the release (titles and bibliographic records as their providers publish them, quoted
excerpts, verbatim surfaces), is not licensed by us and remains its owners'. Quotes are short excerpts, each in the row of the document it comes from, and limited per document: its distinct quoted text stays within 6,000 characters and 25% of its converted text, and quotes beyond one short one per annotation within 2,000 characters and 15%; where the abstract is withheld, everything the row takes from the document, its title included, repeats at most 30% of it (runs of 4 words). Titles, authors, affiliations and surfaces are not counted in the per-document limit (FORMAT.md, "Quotes"). The limits are our own controls on how much of a work the release quotes, not a legal assessment. The release checker in
tools/ is MIT-licensed, as in the data repository.
See LICENSE. Releases are dated and planned weekly; each is a revision of this repository tagged release-<date>.
Personal data. The rows carry authors' names, ORCID iDs and affiliations as the works and their indexes publish
them, the identifiers and links of each work's records in those indexes and forums (which can lead to an author's
profile there), and affiliations and organisations our models and matchers inferred from the text, for one purpose: a
research index of technical AI-safety literature, in which each work is found and attributed by its authors.
We remove the email addresses we detect in what this release ships (each becomes [email]; a link that carries one is left out); detection is automatic and may miss some. To have a record corrected, or to object to or ask for the removal of personal data about you, write to
info@thesafetystack.ai or use the issue forms above.
An internal variant with abstracts and full texts exists for the maintainers and is not distributed.
Team: Tomáš Gavenčiak (lead; Alignment of Complex Systems group, Charles University, Prague), Jacob Livingston Slosser (Sapien Institute; Pioneer Centre for AI, University of Copenhagen), Dan Elton (Metascience Observatory), Gavin Leech (advisor). Cite as: The Safety Stack dataset, release 2026-11-20, https://huggingface.co/datasets/thesafetystack/the-safety-stack.
Annotation quality
How far this release's annotations (production: gpt-6-luna) agree with three language-model judges, Claude Opus 5.5,
GPT Sol 6.1 and Claude Sonnet 5.5, who labelled samples of build ws1qpn8myzxah blind, reading the whole document;
the numbers compare their labels with this release's build wz1j5zcmaqd3b on the same documents (a document no longer
in the build, or whose input changed since, is not scored). 57 of the 600 sampled documents are not scored (50 its
input changed since the judges read it; 4 no longer in the judged population; 3 no longer in the build). The reference
is the label at least 2 of the 3 judges give; for technique presence production is compared with Opus and with Sol
separately. Intervals are 95 % document-bootstrap intervals; the triage row is a population estimate from a weighted
sample of the records triage judged, weighted by this build's population. The stand-in annotations are not in the
samples.
| Judgement | Reference | n | Precision | Recall | F1 |
|---|---|---|---|---|---|
| Triage: fetch the full text (relevant or borderline vs not relevant), from a record's title and metadata | majority of the 3 judges | 155 of 157 records, weighted | 0.65 [0.52, 0.78] | 0.91 [0.82, 1.00] | 0.76 [0.65, 0.85] |
| Screen: keep (relevant or borderline vs irrelevant), from the text | majority of the 3 judges | 99 documents | 0.91 [0.83, 0.98] | 0.93 [0.85, 0.98] | 0.92 [0.87, 0.97] |
| Affiliations: (author, affiliation as written) pairs | majority of the 3 judges | 94 documents | 0.82 [0.68, 0.95] | 0.84 [0.69, 0.97] | 0.83 [0.69, 0.96] |
| Models: (model, role) pairs at strength 2 or more of 4 | majority of the 3 judges | 95 documents | 0.90 [0.85, 0.94] | 0.86 [0.81, 0.90] | 0.88 [0.84, 0.91] |
| Techniques: presence (taxonomy slug, any role) | Claude Opus 5.5 | 98 documents | 0.78 [0.73, 0.82] | 0.69 [0.64, 0.73] | 0.73 [0.69, 0.77] |
| Techniques: presence (taxonomy slug, any role) | GPT Sol 6.1 | 98 documents | 0.88 [0.84, 0.92] | 0.55 [0.50, 0.60] | 0.68 [0.64, 0.72] |
Relevant documents triage drops: of the 153,173 eligible records triage marked not relevant, whose full text was therefore not fetched, none of the 104 sampled ones is relevant by the majority, so the estimate is 0; the Wilson 95 % upper bounds in its confidence strata are 1,513 (confidence up to 0.94), 6,010 (confidence 0.95-0.98), 9,973 (confidence 0.99).
- The judges are language models, not human annotators.
- How far the judges agree with each other bounds what these numbers can show (technique presence F1 between Opus and Sol: 0.72).
The judges' labels, the samples, the method and the full tables for build ws1qpn8myzxah are in
evals/judge-2026-10-25j/ of the pipeline repository; release.json (quality.judge_eval) has this release's
numbers and coverage per set.
Coverage
| Document type (providers') | Documents |
|---|---|
| preprint | 19,253 |
| forum_post | 12,809 |
| paper | 9,300 |
| lab_blog | 1,818 |
| conference | 107 |
| none given (NULL) | 72 |
| model_card | 64 |
| system_card | 58 |
| journal | 53 |
| framework_policy | 15 |
| book | 6 |
| other | 3 |
Published 2004-2026 (632 undated). Sources (the anchor record's channel): openalex 16,602, semantic_scholar 10,929, lesswrong 10,431, alignmentforum 2,202, arxiv 1,134, openai_blog 523, miri_blog 341, uk_aisi_blog 144, redwood_blog 109, far_ai_blog 96, deepmind_blog 95, metr_blog 77, and 31 more channels. Document languages (model-assigned): en 43,528, zh 7, de 5, ru 5, ja 3, ko 3, es 2, fr 2, pl 2, pt 1.
Affiliations: 139,114, of which 130,088 (93.5 %) name at least one organisation of taxonomies/orgs-0.2.2.yaml (org_ids); by org_status: matched 130,082, ambiguous 578, non_org 649, generic 713, unmatched 7,092.
Files
| Path | Rows | What |
|---|---|---|
data/corpus.parquet |
43,558 | The full row per document: bibliographic fields, identity, content identifiers, release facts and every annotation (nested). The contract and the hub's only configuration; everything else resolves its ids or describes it. |
config/configs.json |
29 | One entry per config_id: the call that produced annotations (stage, prompt name and digests, model, parameters, taxonomy pin) or the overrides that supplied them. |
config/overrides.jsonl |
624 | The active overrides of the pipeline's output (maintainers' decisions and stand-in annotations), one line each: override_id, run_id, kind, target, operation, key, field, value, note, author, method, created_at. |
taxonomies/ |
The techniques taxonomy and the models catalogue the annotations were graded against (byte-identical to the pinned versions; one file per version: annotations kept from an earlier config ship their version beside the current one), and the organisation vocabulary affiliations' org_ids resolve in (byte-identical to the versions the annotations pinned; sha256 of each in release.json, which lists every version per name, oldest first). | |
data/corpus.schema.json |
JSON Schema (2020-12) of one row of data/corpus.parquet, self-contained: every kind's record and payload inline; every node's exact Arrow type as x-arrow; closed lists as enum, open lists as x-open with their known values in x-values. | |
release.schema.json |
JSON Schema of release.json: strict on the release's identity, format version, taxonomy pins and file list; the diagnostic sections are free objects. | |
config/configs.schema.json |
JSON Schema of config/configs.json (an array of configs; the JSON-encoded fields stay strings). | |
config/overrides.schema.json |
JSON Schema of one line of config/overrides.jsonl: the fields and the closed list of kinds; the value per kind is described in FORMAT.md. | |
tools/release-validator.py |
The release checker (uv run tools/release-validator.py . --rows all), byte-identical to the taxonomies repository's: the format version, every file against release.json and its schema, the ids, the taxonomy pins. |
|
VOCABULARY.md |
A sentence for every value word that occurs in this release, and a glossary. | |
FORMAT.md |
The file format: every field, the ids and where they resolve, the rules and the recipes. | |
README.md |
The dataset card, with the annotation quality table. | |
LICENSE |
CC0 1.0 for our annotations and documentation, with the third-party carve-outs. | |
release.json |
The manifest: every other file with its size and sha256, row counts, the build, the policy; also the status remap counts (remap), the overrides chain and how many records it changed per internal workflow (overrides.applied), the stand-ins, the affiliations by org_status (orgs; no_record = the document has no affiliation_orgs record), per annotate stage what it ships (stages: its selection digest, its target config and the earlier configs it accepted, each with config_id and the documents under it), and parent_release (null in this release). |
Release record
- Release
2026-11-20, variantpublic, created 2026-11-20T03:54:08Z. - Built from pipeline build
wz1j5zcmaqd3b(planbatch-2026-10a, corpusr0st9nrghbyjd.corpus) by release code071a446f6e4d+image. - Rows: 80,846 documents with text in the build, 43,558 screened relevant and shipped.
- Policy: abstracts
withheld_pending_rights_review; textsnot_distributed; quote cap 300; quotes limited per document (policy version 2); detected email addresses removed. - Overrides: 624 active rows in
config/overrides.jsonl. release.jsonis the full record (every other file with size and sha256).
Fields of data/corpus.parquet (top levels; FORMAT.md has every field)
| Field | Type | Nullable | Description |
|---|---|---|---|
release_id |
string | no | Release this row comes from (the export's UTC date, YYYY-MM-DD). Constant within a file; keeps rows self-identifying when releases are stacked. Also in the parquet file metadata and release.json. |
doc_id |
string | no | Document id (13 lowercase alphanumerics). Stable across releases; a merged document keeps the surviving id and lists the others in identity.former_doc_ids. |
title |
string | yes | Title (merged; field_sources says which source supplied it, origin human where the maintainers fixed it, an override of kind metadata_fix in config/overrides.jsonl; identity.members[].title keeps each source's). |
abstract |
string | yes | Abstract or opening excerpt as a source supplied it. NULL in the public variant (release.abstract = withheld_pending_rights_review) and where none exists (none). Third-party text. |
authors[] |
list | yes | Byline authors in order, from the supplying source. |
published_date |
date | yes | Publication date (the merged metadata.published_at, a UTC timestamp internally, as a date). See published_precision; the supplying source is field_sources[field = published_at]; per-source dates are identity.members[].published_at. |
published_precision |
string | no | How precise published_date is: unknown = no date; year_inferred = the date is exactly midnight UTC on 1 January, the providers' placeholder for a year-only date (a genuine midnight-1-January date is marked the same); day otherwise. Values: day, year_inferred, unknown. |
doc_type |
string | yes | Document type from the providers and collectors (the merged metadata.doc_type: the form and venue, e.g. an arXiv record is a preprint whatever it contains), or a maintainers' fix (field_sources origin human). NULL where the collector gave none (hand-added and newsletter-linked URLs and papers). Two model-assigned types also exist: annotations.doc_class.payload.doc_type (the screen call; filled for every screened row) and doc_kind of safety_techniques and model_roles (the genre each annotator read the document as); FORMAT.md says which to use for what, VOCABULARY.md lists all three. Values: preprint, paper, conference, journal, forum_post, lab_blog, system_card, model_card, framework_policy, book, other (an open list: other values may appear). |
is_withdrawn |
bool | no | A source reports the work withdrawn. |
canonical_url |
string | yes | The anchor source's canonical URL (landing page). Fetch-only query parameters (a contact address, a URL signature) are removed. |
field_sources[] |
list | no | Provenance of each merged bibliographic field (title, abstract, authors, published date, doc_type). |
identity |
struct | no | Which source records make up this document and how they were joined. |
content |
struct | yes | The document text the annotations were made on, as two byte identifiers (the fetched source file and our Markdown conversion of it) plus where it came from. release.text says whether this variant carries the converted text itself (content.text.body). |
release |
struct | no | What this release did to this row. Lets a reader tell a withheld value from an absent one, a cut quote from a short one, and see how much of the document is quoted. |
annotations |
struct | no | Every annotation of the document: one column per kind, each a struct {metadata, payload}; NULL = the kind was not run for this document (affiliation runs only on papers, preprints, conference papers and journal articles whose text came from a PDF, arXiv HTML or publisher XML). metadata has the same type in every kind. Kinds produced by one call share config_id and decided_at: the screen call yields screen, doc_class, stack_relevance, shallow_review_relevance and text_usability; the model_roles call yields model_roles and model_name_matches, except where a stand-in replaced model_roles (FORMAT.md). |
- Downloads last month
- 99