Public-release layout: train/validation/calibration/test splits, nested metadata, dataset card

README.md

@@ -1,130 +1,195 @@
 ---
-license: cc-by-4.0
-task_categories: [text-classification]
-language: [en]
-size_categories: [1M<n<10M]
+pretty_name: IdeaLens-1M
+license: odc-by
+language:
+- en
+task_categories:
+- text-classification
+tags:
+- ai-text-detection
+- idea-provenance
+- outlines
+size_categories:
+- 1M<n<10M
 configs:
 - config_name: default
   data_files:
   - split: train
     path: data/train-*.parquet
+  - split: validation
+    path: data/validation-*.parquet
+  - split: calibration
+    path: data/calibration-*.parquet
+  - split: test
+    path: data/test-*.parquet
 ---
 
-# CC-2026 De-leaked Outlines (1M)
+# IdeaLens-1M
 
-1,001,145 English documents paired with **role-labelled outlines** for idea-level AI-text
-detection. Each row holds the source document, an outline extracted from it, and a
-*de-leaked* outline in which surface style has been paraphrased away so only the ideas
-and their structural roles remain. The intended use is detecting AI authorship from
-**what a document says and how it is organised**, rather than from its prose style.
+IdeaLens-1M is the training corpus of [IdeaLens](https://huggingface.co/rishanthrajendhran/IdeaLens), a detector of
+**who came up with the ideas** in a document, rather than who wrote its words. It pairs 1,001,145 English web
+documents with **role-labelled outlines** of their ideas: an ordered list of items, each giving one idea and the
+role it plays in the document (for example *Central Development*, *Source Viewpoint*, *Open Question*).
 
-**The `source` label is the Pangram prediction.** Documents were selected so that the intended label matches the detector's call, so `source` and `pangram_prediction` agree on 100.00% of rows. A model trained here is learning to reproduce Pangram's judgement, not a human-annotated ground truth.
+Every document carries a human or AI label. Each outline comes in two versions: the outline as extracted, which is
+what IdeaLens reads when scoring a new document, and a paraphrased outline with the document's wording removed,
+which is what IdeaLens was trained on.
 
-## Splits
+## Quick start
+
+```python
+from datasets import load_dataset
 
-The `split` column carries the assignment; all rows ship in one physical `train` config.
+train = load_dataset("rishanthrajendhran/IdeaLens-1M", split="train")
+row = train[0]
+print(row["label"], row["format"])
+for item in row["paraphrased_outline"]["items"]:
+    print(f"[{item['role']}] {item['content']}")
+```
 
-| split | rows | % human | purpose |
-|---|---:|---:|---|
-| `train` | 842,301 | 44.03% | model fitting |
-| `val` | 29,975 | 83.35% | checkpoint selection |
-| `calibration` | 80,000 | 100.00% | per-format thresholds |
-| `test` | 48,869 | 50.29% | held-out evaluation |
+IdeaLens reads an outline as one line per item, `[role] content`, joined with newlines.
 
-`calibration` is humans only and disjoint from `val`, so a threshold is never read off
-the documents used to pick the checkpoint. It holds 10,000 humans per format, putting
-50 documents below a 0.5% quantile — a realised-FPR 95% interval of 0.371%–0.647%.
-**Per-format thresholds are required**: the 1%-FPR cut varies ~16× across formats, so a
-single global threshold realises anywhere from 0.16% to 1.66% FPR depending on format.
+## Splits
 
-`test` is balanced by format (49.5–52.0% human) and by length (49.1–52.6% human).
-`val` is deliberately human-heavy; use AUC or balanced metrics on it, not raw accuracy.
+| Split | Documents | Human | AI | Purpose |
+|---|---:|---:|---:|---|
+| `train` | 842,301 | 370,882 | 471,419 | model training |
+| `validation` | 29,975 | 24,985 | 4,990 | checkpoint selection |
+| `calibration` | 80,000 | 80,000 | 0 | fitting decision thresholds |
+| `test` | 48,869 | 24,574 | 24,295 | held-out evaluation |
 
-## Schema
+`calibration` contains only human documents, 10,000 per format, and is disjoint from `validation`. The IdeaLens
+decision thresholds are the score levels that flag 1% of these documents, either over all of them (global cut) or
+within each format (per-format cut). `validation` is mostly human, so report AUC or balanced metrics on it rather
+than accuracy. `test` is close to balanced within every format.
 
-### Identity and labels
+## Fields
 
-| column | type | notes |
+| Field | Type | Description |
 |---|---|---|
-| `id` | string | unique document id |
-| `source` | string | **the label** — `human` or `ai`. AI is the positive class. Named `source` for continuity with the 391k release |
-| `split` | string | `train` / `val` / `calibration` / `test` |
+| `id` | string | Document id: the FineWeb record id (`<urn:uuid:...>`) for the 903,470 pages from the 2024 and 2025 crawls, a 64-character hash for the 97,675 pages from the 2026 crawls |
+| `text` | string | The full document |
+| `label` | string | `human` or `ai` (see *Labels*) |
+| `format` | string | WebOrganizer format, one of eight: Nonfiction Writing, Knowledge Article, Personal Blog, News Article, Academic Writing, User Reviews, Personal About Page, Creative Writing |
+| `topic` | string | WebOrganizer topic (23 values) |
+| `url` | string | Source URL |
+| `date` | string | Crawl timestamp (ISO 8601) |
+| `word_count` | int | Words in `text` (501 to 17,809) |
+| `outline` | struct | The outline as extracted from the document |
+| `paraphrased_outline` | struct | The same outline, paraphrased to remove the document's wording |
+| `extractor` | string | Model that extracted the outline: `gemini-3.7-flash` or `gemini-3.1-pro` |
+| `paraphraser` | string | Model that paraphrased it: `gemini-3.1-pro` |
+| `metadata` | struct | Detector outputs available for the document (see *Metadata*) |
+
+Both outline fields share one structure:
 
-### Document metadata
+```
+document_description: string          # what kind of document this is: genre, audience, register
+global_themes:        list<string>    # the document's main points
+items: list<{
+    role:     string,                 # the role this item plays, from the format's role vocabulary
+    content:  string,                 # one idea, stated in one or two sentences
+    verbatim: bool                    # true if the content quotes the document rather than summarising it
+}>
+```
 
-| column | type | notes |
-|---|---|---|
-| `format` | string | 8 values: Nonfiction Writing, Knowledge Article, Personal Blog, News Article, Academic Writing, User Reviews, Personal About Page, Creative Writing |
-| `role_format` | string | role taxonomy used when labelling outline items; same 8 values as `format` |
-| `topic` | string | 23 values (Finance & Business, Health, Sports & Fitness, …) |
-| `url` | string | origin URL |
-| `date` | string | source date |
-| `word_count` | int64 | 501–17,809. The corpus floor is 501 words |
-| `token_count` | int64 | 529–132,083 |
+The paraphrase keeps every item's role, order and content, and changes only the wording.
+
+## How the corpus was built
+
+- **Documents and labels.** Documents are English Common Crawl pages from the 2024, 2025 and 2026 crawls, drawn from
+  the WildAI corpus (citation below), which builds on FineWeb and annotates each page with a WebOrganizer format and topic and a Pangram authorship prediction. We kept documents of
+  at least 500 words, in eight long-form formats, that Pangram labels as entirely human-written or entirely
+  AI-generated. Documents Pangram labels as mixed or AI-assisted were left out.
+- **Outline extraction.** For each format, a role vocabulary lists the recurring functions an idea can serve in that
+  kind of document. An LLM, given the document, its format's role vocabulary and six worked examples, writes the
+  outline. Most documents were extracted by `gemini-3.7-flash`; part of the corpus, from an earlier build, by
+  `gemini-3.1-pro`. The `extractor` field records which.
+- **Paraphrasing.** `gemini-3.1-pro` rewrote every outline item without seeing the source document, so that the
+  document's own wording does not carry into training.
+
+The prompts, role vocabularies and worked examples are in the code repository: *link added on publication*.
+
+## Labels
 
-### Outlines
+The labels come from the data source, which assigned them with the Pangram AI-text detector: Pangram 3.3.2 for pages
+from the 2024 and 2025 crawls, and Pangram 4 for pages from the 2026 crawls (`metadata.pangram.model` records which).
+They are **silver labels**: they record whether Pangram judged the prose to be human-written or AI-generated, not an
+annotation of who conceived the ideas. IdeaLens is trained on these prose labels through outlines whose wording has
+been removed, which pushes it to learn from the ideas themselves.
 
-`extracted_outline` (stage 1) and `deleaked_outline` (stage 2) share one struct:
+## Metadata
+
+`metadata` holds the detector outputs that exist for a document. Each block is either complete or null:
 
 ```
-struct<
-  document_description: string,
-  global_themes:        list<string>,
-  items: list<struct<
-    id:        int64,
-    content:   string,
-    verbatim:  bool,     # content lifted from the document rather than abstracted
-    role_name: string    # the item's structural role
-  >>
->
+metadata:
+  pangram:                           # every document
+    prediction:       "Human" | "AI"  # the verdict the label was taken from
+    model:            "pangram-3" | "pangram-4"  # the Pangram model behind the label
+    version:          string | null   # "4.0" on the 10,000 documents we scored with Pangram 4 ourselves
+    score_confidence: {score, confidence} | null      # 603,132 documents
+    fractions:        {ai, ai_assisted, human} | null # 97,675 documents (2026 crawls): share of the text in each class
+  editlens:           {score, bucket} | null          # 57,886 documents (2026 crawls), EditLens AI-editing score
 ```
 
-`source_text` (string) holds the full original document.
+| Documents | Crawls | `model` | Pangram outputs present |
+|---:|---|---|---|
+| 603,132 | 2024, 2025 | `pangram-3` | prediction, score and confidence |
+| 300,338 | 2024, 2025 | `pangram-3` | prediction only |
+| 97,675 | 2026 | `pangram-4` | prediction and fractions (EditLens score for 57,886) |
 
-| column | type | notes |
-|---|---|---|
-| `extractor` | string | stage-1 model. `gemini-3.7-flash` for the 613,132 newly extracted rows; `Gemini 3.7 Flash (6-shot)` and `Gemini 3.1 Pro (6-shot)` for the 388,013 carried over. **The corpus is deliberately mixed-extractor** — the differing labels are meaningful and are what an extractor-invariance test keys on |
-| `paraphraser` | string | stage-2 model: `gemini-3.1-pro-preview` (new) or `Gemini 3.1 Pro` (carried over) |
+## Intended use
 
-### Pangram detector scores
+Training and evaluating detectors of idea provenance, and studying how human and AI-generated documents differ in
+their ideas and structure. A detector trained on this corpus estimates the provenance of a document's ideas; it
+should not be the sole basis for decisions about a person's work.
 
-The two Pangram generations populate **different fields**, so these nulls are
-structural rather than missing data:
+## Limitations
 
-| column | type | populated for |
-|---|---|---|
-| `pangram_model` | string | all rows — `pangram-3` (603,132) or `pangram-4` (398,013) |
-| `pangram_version` | string | 10,000 rows |
-| `pangram_prediction` | string | all rows — `AI` / `Human` |
-| `pangram_score` | double | **pangram-3 rows only** |
-| `pangram_confidence` | double | **pangram-3 rows only** |
-| `pangram_fraction_ai` | double | **pangram-4 rows only** |
-| `pangram_fraction_ai_assisted` | double | pangram-4 rows only |
-| `pangram_fraction_human` | double | pangram-4 rows only |
-| `pangram4_prediction` | string | pangram-4 rows only — kept so consumers of the 391k release read exactly what they read before |
-| `pangram4_fraction_ai` | double | pangram-4 rows only |
-| `pangram4_fraction_ai_assisted` | double | pangram-4 rows only |
-| `pangram4_fraction_human` | double | pangram-4 rows only |
-| `editlens_bucket` | int64 | 5.8% — editlens-derived rows only |
-| `editlens_score` | double | 5.8% — editlens-derived rows only |
-
-`pangram_*` are the generic columns carrying every row's score whatever produced it;
-`pangram4_*` are the original narrow columns. `editlens_*` are a property of one source,
-not of the document, so they are null elsewhere rather than imputed.
-
-## Provenance
-
-Outlines were produced in two stages: extraction with `gemini-3.7-flash` (6-shot,
-thinking level HIGH) and de-leaking with `gemini-3.1-pro-preview` (HIGH).
-613,132 rows were newly extracted for this release; 388,013 are carried over from
-the earlier 391k release with their original outlines.
-
-## Known limitations
-
-- Labels are detector-derived, not human-annotated (see above).
-- The corpus is mixed-extractor by design; models are expected to be invariant to it.
-- Creative Writing and Personal About Page are human-majority because their AI-side
-  pools were exhausted — no allocation could balance them further.
-- Format carries a small amount of label information: a format-only classifier beats
-  the majority baseline by **0.8 points** (mutual information 0.0038 bits, 0.38% of
-  label entropy). Length carries 0.0000 bits and topic 0.0011 bits.
+- English only; documents of at least 500 words; eight long-form web formats.
+- Labels are detector-derived (see *Labels*).
+- Two extraction models were used; detectors trained here are expected to be robust to the extractor, and the paper
+  tests this.
+- Creative Writing and Personal About Page lean human (few AI-labelled documents were available in these formats).
+- Outlines were written by LLMs and can contain extraction errors.
+
+## License
+
+The dataset is released under the Open Data Commons Attribution License (ODC-By 1.0), following FineWeb. The
+document text remains subject to the terms of its original sources and of Common Crawl. The outlines were generated
+with Google Gemini models.
+
+## Citation
+
+```bibtex
+@misc{idealens2026,
+  title  = {IdeaLens},
+  author = {Anonymous},
+  year   = {2026},
+  note   = {Citation added on publication}
+}
+```
+
+Please also cite the source of the documents and labels:
+
+```bibtex
+@misc{wildai2026,
+  title  = {Scaling Laws for Wild {AI} Text in Pretraining Data},
+  author = {Anonymous},
+  year   = {2026},
+  note   = {Citation added on publication}
+}
+@inproceedings{penedo2024fineweb,
+  title     = {The {FineWeb} Datasets: Decanting the Web for the Finest Text Data at Scale},
+  author    = {Guilherme Penedo and Hynek Kydl{\'\i}{\v{c}}ek and Loubna Ben allal and Anton Lozhkov and Margaret Mitchell and Colin Raffel and Leandro Von Werra and Thomas Wolf},
+  booktitle = {The Thirty-eighth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
+  year      = {2024}
+}
+@inproceedings{wettig2025organize,
+  title     = {Organize the Web: Constructing Domains Enhances Pre-Training Data Curation},
+  author    = {Alexander Wettig and Kyle Lo and Sewon Min and Hannaneh Hajishirzi and Danqi Chen and Luca Soldaini},
+  booktitle = {Forty-second International Conference on Machine Learning},
+  year      = {2025}
+}
+```

data/{train-00000.parquet → calibration-00000-of-00001.parquet}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c6b820677146c60e8550e583107fa721e90c7aa1dd2cf26d6d97985f38d857ed
-size 471120811
+oid sha256:89d9d2a6db68df2e53a85809a2d00a4ba55f63206382ca7d7cefd1afb702634d
+size 482608527

data/{train-00001.parquet → test-00000-of-00001.parquet}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:19a2197aeee22c8089a9e91644ffe16ba7afce2531ddb944f6ce626a65f860e8
-size 470396010
+oid sha256:2be3bbf03934b58a9f911c019526c90fcf7f6c445cbfbcffd1dff30e6a8da692
+size 378648552

data/{train-00002.parquet → train-00000-of-00011.parquet}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b66e2f468f0ebc19756865604f4b758e3e76654c27e48ee09d83826c6ae6cde7
-size 469847444
+oid sha256:8ae39646e20d2fa8e2136766479902f6358d765dafd32a4ef4f0d6af08f96810
+size 465965041

data/{train-00003.parquet → train-00001-of-00011.parquet}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:aea72c54d77a9d4481509cba62a0bda333391ea2d062efca3ae2d3e2c6d37e94
-size 470859665
+oid sha256:32fd3aa76c75e7810d7c83cae91dda61047352de9d47fad6790222d6300f46fa
+size 465075609

data/train-00002-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3efd96bd1eec60bc1b4f69403b057c6e96ddd61035caae30a4d05e0563c05fde
+size 465894384

data/train-00003-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a81b83f41e371f30d54211d42962393045182d30ba925923d5577daeb20d4ae
+size 465401062

data/train-00004-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dcce42df3dd3ab8def5cd691bf508350c06a1db8e31bb5fefbf9c54d6f01271a
+size 464602097

data/train-00004.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:136f2809f3374b80596e025ab2b157d92f06719f254e452ae64af1c53f3209f7
-size 470089586

data/train-00005-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aab1d6ca519328205e0eaaad36d403fb63938d5a44d801fe9a059e699330005b
+size 465705493

data/train-00005.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:aeac7ea263c92355175a1a43ee2059b4268e92f1a664edb58c31334329c755b9
-size 470499265

data/train-00006-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8519efcf10ff6a9a175f9b8a971edeba01fa23fdecd3c10c77fd78103a1ee5a8
+size 466076807

data/train-00006.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:bfc0ce07fc6337d663d30418890fb8c858fb216484f3f6c0f77da6043ebdf732
-size 470395874

data/train-00007-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:734d7c63eac2b00d4415440a484b7873e83be719e1f87f20b857f4ed84fe0cdf
+size 454537514

data/train-00007.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:05c799fffa0a997f00d9a6a5e43502497d8fb25c16ee142702af27cf28d01d92
-size 313000801

data/train-00008-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fa673583b86b7e4ec8a0f473ee096b5be879b011f4d5299497546d107685eef
+size 455892247

data/train-00008.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:adb3dde68ff8f4f9a69fe2c77d35f1453a70d86c0e2897b50a61f36cec11ec08
-size 470771439

data/train-00009-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc2ba016787d9f083994d03348239d73d449bc0dfaeed0d689b82ca734d165fd
+size 482013917

data/train-00009.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:3e3592d50a704d3477e6a6e9cee51bcd2ca36eb0378bc3e9ae5f7cb36ee36416
-size 445662574

data/train-00010-of-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e30b9eb33ecd5df0f5e6b23d0a026712185011138142abe0d809a9d5eb2b940
+size 254679388

data/train-00010.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:e3eaa28a6a6f31ae1d3589787f87307fe3b7d03fc32f61e590fa7391d2d216e9
-size 490825673

data/train-00011.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:3103223106dbea4d63204a3486523cf7e5b8e4118630974f6e0c774590f9bdf6
-size 496079895

data/train-00012.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:41b5d795152feeeb147fd790e6a9115883f0305909b001e5eca21b32042790dd
-size 422122091

data/validation-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0a998bfcb4a153c4ca1f439567521e4ee6d301cfcef954a322438d11f9e4dff
+size 184177699