open-text-detector/heterogeneous-ai-spans

Heterogeneous AI Spans

4400 mixed documents · 4400 matched human controls · 5283 AI replacement spans · 4400 distinct source passages

An English dataset for locating AI-written sections inside otherwise human-origin documents, with exact character spans and the identity of the model that wrote each replacement. The labels are construction provenance: they record which text was retained and which text was generated, rather than the judgment of another AI detector.

Human means non-AI. Retained source text has documented historical provenance and remains explicitly marked as a human-origin candidate, not independently certified wording. Every source has human_verified=false; the richer JSONL also retains authorship="human_candidate" and binary_target=null.

Version 1.4.0: formatting correction

No new generations. The same 4,400 mixed documents and 4,400 controls retain their IDs, source groups, splits and model attribution. Text, source text, span offsets, text hashes and character fractions are corrected consistently in every configuration.

The default view preserves genuine paragraphs, including blank lines containing spaces or tabs. CRLF/lone CR are normalized, physical hard wraps are joined, whitespace runs are standardized, paired _italics_ / Markdown emphasis markers are removed, curly quotes become straight quotes, Unicode en/em dashes become --, and Unicode hyphens and ellipses become ASCII forms. Genuine diacritics and mathematical symbols remain; broad NFKC normalization is deliberately disabled. The same deterministic rules apply to both authorship classes.

Paragraph length and paragraph-edge cues remain in the default view. Use mixed_layout_neutral, human_controls_layout_neutral or all_layout_neutral to collapse all whitespace to spaces. These contain the same samples, with their own remapped offsets, and must not be combined with the default views as additional training examples. Apply the same normalization policy to evaluation and inference inputs, retaining offset maps if predictions must refer to raw text.

neutral = load_dataset("open-text-detector/heterogeneous-ai-spans", "mixed_layout_neutral")

Formatting audit compares the original and both corrected views. Raw full records preserve v1.3.0 byte-for-byte. Default full records and layout-neutral full records use corrected text and offsets. generation.raw_text retains the actual model response, while generation.text / generated_text contain the derived training text and generation.postprocessing identifies the policy. Original prompts and model provenance are unchanged. Upstream parent ranges and recorded generation-quality measurements still refer to raw text; inherited audits apply to the raw records. Earlier preservation statements below describe the historical releases.

New offsets are incompatible with the old text. Join samples across views by document ID, and always take text and spans from the same view. The correction removes formatting conventions; it does not equalize dialogue, content, diacritics or punctuation frequency, and does not establish detector accuracy after retraining.

Start here

from datasets import load_dataset

data = load_dataset("open-text-detector/heterogeneous-ai-spans", "mixed")
row = data["train"][0]
for span in row["spans"]:
    fragment = row["text"][span["start"]:span["end"]]
    print(span["label"], span["reported_model"] or span["requested_model"], fragment[:120])

If this repository is private, sign in with hf auth login or provide your existing read token to the library. Do not place tokens in source files.

• Training / evaluation: use the Parquet configurations below.
• Full provenance: records/all.jsonl includes original source metadata, briefs, exact prompts, model responses, quality measurements and author metadata.
• Visual inspection: download the complete ZIP, extract it, and open review/index.html. The review works offline, with model, cohort and split filters; retained human text is green and AI text is orange.
• Format details: schema guide and full-record JSON Schema.
• Verification: release manifest, split counts, quality reports, and SHA-256 checksums.

What's included

Writer	original	ml_papers	expansion	gpt_general	gpt_ml_papers	overnight_batch_01	overnight_batch_02	overnight_batch_03	Mixed total	Recorded identity
Haiku	100	100	100	0	0	200	200	200	900	claude-haiku-4-5-20251001
Sonnet	100	100	100	0	0	200	200	200	900	claude-sonnet-5-5
Current Opus	100	100	100	0	0	200	200	200	900	claude-opus-5-5
Opus 3	100	0	0	0	0	0	0	0	100	Browser UI label Opus 3
GPT-6.1 Sol	0	0	0	100	100	200	200	200	800	Requested gpt-6.1-sol; served identity unreported
GPT-6 Luna	0	0	0	100	100	200	200	200	800	Requested gpt-6-luna; served identity unreported

These are the identities recorded by the generation backends. Requested and reported identities remain separate. Opus 3 used a visibly selected browser model; its API checkpoint, revision and sampling parameters were not reported. It must not be silently equated with a particular dated API checkpoint. Claude models ran through Claude Code subscription authentication; GPT writers ran through Codex subscription authentication, retaining requested-only attribution when served identity was unreported; Opus 3 ran through the browser UI. No OpenRouter/open-weight generations are included.

Each source passage is assigned to one writer; different models did not rewrite the same source passage. Some long parent books or reports contribute multiple disjoint excerpts. The original batch includes fiction, encyclopedia/reference text, and professional economic reports. The ML subset contains narrative excerpts, not full papers. Its current-Opus examples had already completed before the decision to focus future Opus runs on stories.

Version 1.1.0 expansion

The added batch contains 100 new documents each from Haiku, Sonnet and current Opus, plus 300 matched controls, with no additional Opus 3 generation. Opus uses historical-fiction passages only. Haiku and Sonnet each receive 15 Gutenberg passages, 35 Standard Ebooks passages, 25 WikiText passages and 25 Beige Book passages. New excerpt IDs, normalized texts and parent ranges are checked against the previous release. Existing book/author/report split assignments are inherited; the older 1,400 document IDs and text are retained.

The redundant source_text_sha256 Parquet column has been removed from every configuration, including the previous data. Compute it from source_text when needed. The full provenance JSONL retains the recorded upstream hashes. The previous v1.0.0 Hub commit remains available as version history.

Version 1.2.0 GPT expansion

Adds 100 general documents and 100 distinct pre-2015 ML-paper excerpts for each of GPT-6.1 Sol and GPT-6 Luna: 400 mixed documents plus 400 controls. General passages per writer comprise 10 Gutenberg, 40 Standard Ebooks, 3 WikiText and 47 Beige Book excerpts, reflecting the remaining unused eligible source pool. The 200 additional papers exclude every previously used paper.

GPT writers used Codex subscription authentication with requested IDs gpt-6.1-sol and gpt-6-luna, and explicit low reasoning effort. The CLI did not report the served model: reported_model=null, model_identity_status="requested_only". No served checkpoint is inferred. Briefs still use Haiku, as in previous cohorts. New paper excerpts require at least five continuous clean prose paragraphs, with at most 65% of source characters replaced; the same 3/4/6-paragraph blocks, brief bottleneck and 15% source-copy ceiling apply.

The v1.2.0 release preserved its 2,000 earlier full records exactly, including IDs, source/generated text, prompts and split assignments. ml_papers_mixed includes the earlier and GPT paper cohorts; gpt_ml_papers_mixed isolates these 200 papers.

Version 1.3.0: completed overnight batches

Adds 3,000 mixed documents and 3,000 matched controls from 3 completed, audited batches. Each of Haiku, Sonnet, current Opus, GPT-6.1 Sol and GPT-6 Luna contributes 600 new documents; no new Opus 3 examples are included. The additions comprise 1,200 pre-2015 ML excerpts and 1,800 general passages. General passages are predominantly historical fiction, supplemented by two historical Hansard excerpts. Current Opus uses fiction only; the other four writers each receive equal ML/general quotas.

All 2,800 full records from v1.2.0 are preserved exactly. overnight_mixed isolates these additions; individual overnight_batch_01_mixed etc. configurations isolate each parameter profile. Only completed batches are included; the remaining generation campaign is separate from this release.

Cohort	Paragraphs per block	Maximum blocks	Source replacement cap	Brief sentences	Context characters
overnight_batch_01	2, 3, 4	1	55%	2–3	400
overnight_batch_02	2, 3, 5	2	60%	2–3	600
overnight_batch_03	2, 4, 6	2	65%	2–3	900

These are selection limits; actual replacement counts and lengths are in each record. Context uses complete retained paragraphs, so its character budget is soft. ML excerpts contain three continuous eligible paragraphs; multiple disjoint excerpts can come from the same paper. Previously selected paragraph ranges are excluded, and paper/author groups retain one split across old and new cohorts. New books are checked against the archive's title/author header and pinned raw bytes.

Three complete mixed documents per writer per new batch were read (45 documents); six rendered original PDF pages were compared with the corresponding excerpts. Mechanical span/provenance/copy audits cover every released replacement. The et al. sentence-counting bug was corrected during the third batch; accepted prompts and model responses remain in the full records.

Hugging Face configurations and splits

mixed is the default. All configurations use the same documented Parquet schema and preserve their original split assignments.

Configuration	Train	Validation	Test	Total
mixed	3509	297	594	4400
human_controls	3509	297	594	4400
all	7018	594	1188	8800
original_mixed	339	30	31	400
ml_papers_mixed	1392	151	157	1700
expansion_mixed	233	33	34	300
gpt_general_mixed	139	19	42	200
gpt_ml_papers_mixed	166	17	17	200
overnight_batch_01_mixed	787	70	143	1000
overnight_batch_02_mixed	806	47	147	1000
overnight_batch_03_mixed	800	54	146	1000
overnight_mixed	2393	171	436	3000

controls = load_dataset("open-text-detector/heterogeneous-ai-spans", "human_controls")
papers = load_dataset("open-text-detector/heterogeneous-ai-spans", "ml_papers_mixed")
original = load_dataset("open-text-detector/heterogeneous-ai-spans", "original_mixed")
expansion = load_dataset("open-text-detector/heterogeneous-ai-spans", "expansion_mixed")
gpt_general = load_dataset("open-text-detector/heterogeneous-ai-spans", "gpt_general_mixed")
gpt_papers = load_dataset("open-text-detector/heterogeneous-ai-spans", "gpt_ml_papers_mixed")
overnight = load_dataset("open-text-detector/heterogeneous-ai-spans", "overnight_mixed")
everything = load_dataset("open-text-detector/heterogeneous-ai-spans", "all")

For a fixed snapshot, add revision="v1.3.0" to load_dataset. The original release remains available at revision="v1.0.0"; its schema included the now-removed redundant source-hash column.

The configurations overlap; do not concatenate all with the others. Each control shares the original source with one mixed document and is intentionally in the same split. Source IDs, group IDs and normalized duplicate texts are isolated across train/validation/test across all cohorts. JMLR split groups are paper-level; historical fiction uses the recorded book/author grouping. The source pair is a useful unit for analysis; treating paired controls and mixed documents as independent observations can inflate confidence intervals.

Source provenance

Collection	Source passages	Evidence
Project Gutenberg	1948	Named historical works and retained raw-source hashes; current ebook revision dates are not independently established.
Standard Ebooks	350	Named historical works and pinned repository commits from before 2023.
WikiText-2 raw	156	The 2016 Wikipedia-derived release, retained raw Parquet/article hashes and mechanical punctuation restoration. Individual article revision IDs are unavailable.
Federal Reserve Beige Book	244	Dated official narratives and saved source hashes; capture-version dates are not independently certified.
JMLR, 2000–2014	1700	Publication year agrees between the journal index and PDF; original PDF/layout/index hashes, authors and paragraph locations are retained.
Historical UK Hansard	2	Retained parliamentary/source provenance, source date and pinned parent-text/release hashes.

The ML sample was selected with a seeded shuffle from JMLR volumes 1–15. The original paper cohort requires at least six continuous usable prose paragraphs; the GPT paper cohort requires at least five. The new overnight cohorts use three continuous eligible paragraphs per excerpt. Selection does not join separate eligible runs across rejected paragraphs. Extraction used Poppler rather than a language model, with recorded line joining, dehyphenation and normalization. Heuristic paragraphs can include a figure caption joined to nearby narrative, and inline mathematical notation loses some visual layout; source coordinates and the visual-review findings retain these artifacts. Selection therefore favors PDFs with usable narrative extraction; it is not a representative sample of all ML papers. Human-only source material was drawn from the upstream documented collections, not accepted on the strength of a detector score.

The earlier historical-fiction/reference/finance source collection was pinned to upstream repository revision b618ca42b85fcb4cd8ddcc60cf18ed957262cd6a; subsequently acquired sources retain their own acquisition evidence and hashes. The generation plans and selection details are recorded in provenance; original source excerpts and metadata are in sources/originals.jsonl.

How mixed documents were made

1. Earlier cohorts choose 3, 4 or 6 source paragraphs per block; overnight cohorts vary block sizes from 2–6 paragraphs, with one or two replacements and 55–65% source-character caps.
2. Condense each selected block into 2–3 complete sentences, using Haiku. This is a multi-paragraph content brief, not a short list of keywords.
3. Give the assigned writer the brief plus bounded neighboring context, excluding the selected original blocks. Ask it to expand the brief into document prose.
4. Splice the generated text into the original source and construct an exact, contiguous span partition with retained-source and replacement coordinates.
5. Save the full prompt/response provenance and an unchanged control from the same source.

Successful generations were cached. Failed quality attempts were retried; the exact accepted prompt is saved. Corrective prompts for memorized fiction could change fictional names/details or use modern wording. Model-reported metadata and unreported fields are kept distinct. Subscription/UI generation does not establish deterministic replay or disclose every sampling setting.

Quality checks

• All 8800 release records pass the full JSON Schema and exact reconstruction/span checks. Text and prompt content are unchanged by packaging.
• All 5283 AI replacement blocks passed the configured source-copy checks, including checks against the full own-parent text rather than only the excerpt. The check uses normalized exact 8-gram coverage with a 15% ceiling. This is not an Internet-wide plagiarism check and cannot prove that every generated sentence is novel.
• Final audits report no missing planned documents or outstanding prose-review flags.
• Earlier modern-model cohorts reviewed five outputs per participating writer. Each released overnight batch reviewed three complete documents per writer. Five actual PDF-page renders were checked for the original paper cohort, and five additional pages for the GPT paper cohort. Two additional pages were inspected for each of the 3 released overnight batches. The full dataset was not manually reviewed sentence by sentence.
• Two cached Opus 3 passages that substantially copied other chapters of their source books were quarantined and replaced. The final export contains their accepted replacements; the quality history is retained in the original-cohort quarantine report.
• Factual/causal fidelity, exact output length and exact paragraph count are not acceptance requirements. The objective is usable, correctly attributed text for a detector; rewritten paper passages are not reliable accounts of the original research.

Span and author semantics

start and end index the resulting text using Unicode code points and half-open intervals: text[start:end]. They are not byte, token or JavaScript UTF-16 offsets. The spans cover every character without gaps or overlaps, including separators. The offline review converts to code points before slicing.

source_start / source_end refer to the retained original source_text. For a human span, its text is exactly the corresponding source slice. For an AI span, that range identifies the original block being replaced and may have a different length. AI reported_model and requested_model identify the writer; retained human spans carry the source author where known. Unknown human authors remain null.

The Parquet view intentionally provides a stable, compact author representation. The authoritative full record in JSONL contains the richer author object, full usage metadata, generation and summarization provenance, prompts, source hashes and replacement history. Join the two views by id. See the schema guide for an example.

Files and reproducibility

data/<configuration>/{train,validation,test}.parquet
records/{all,mixed,human-controls}.jsonl
sources/originals.jsonl
schema/{README.md,full-record.schema.json,arrow-schema.txt}
audits/<cohort>/
provenance/                    # portable selection/configuration and copy-policy exports
manifest/                      # release, splits, upstream hashes, metadata transformations
review/index.html              # self-contained offline review of all 8800 records
examples/load_and_inspect.py
reproduction/                  # generator, collection/audit scripts, tests and configuration
SHA256SUMS.txt
heterogeneous-ai-spans-v1.4.0.zip

The source text, generated text, exact spans, text hashes and prompts are preserved. Local absolute file paths are converted to portable upstream artifact hints; references that pointed at local corpus files are replaced with the recorded upstream public URL. Private Claude chat URLs, including URL-valued request IDs, are replaced with stable SHA-256 references. These changes are documented in the release manifest. Frozen original plan IDs refer to the original local plans; the portable planning exports are clearly labeled representations, not byte-identical plan files.

Full original PDFs/books, raw browser responses/session state, local logs, credentials and machine-specific runtime files are not included. Full-parent reference hashes and check outcomes remain available. The bundled scripts explain regeneration, but access to the upstream sources and compatible model subscriptions/backends is required; the historic browser model may not remain available.

Intended use and limitations

Suitable for span-level AI/non-AI detection research, mixed-authorship segmentation, matched-pair experiments and exploratory model attribution. Labels describe the construction process and depend on the documented source-origin evidence. This pilot is synthetic, small, English-only and limited to 6 recorded writers; its performance is not a claim about real-world detector reliability or unseen models. Prompt style, source domains, extraction artifacts and replacement boundaries may become shortcuts. Do not use a classifier trained only on this collection to make consequential claims about an individual's authorship.

Rights and citation

The collection combines different source terms. No single permissive content license is asserted. Attribution and the per-source rights descriptions remain in every full record and in the Parquet source_license field. In particular, WikiText's recorded license-version discrepancy and historical JMLR paper-specific license uncertainty remain unresolved. See LICENSE.md before reuse or redistribution; the packaging does not override upstream rights or service terms.

@misc{open_text_detector_heterogeneous_ai_spans_2026,
  title = {Heterogeneous AI Spans: Human/AI Mixed Documents},
  author = {{Open Text Detector}},
  year = {2026},
  url = {https://huggingface.co/datasets/open-text-detector/heterogeneous-ai-spans},
  note = {Version 1.3.0; 4400 mixed documents and 4400 matched controls}
}