The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.

Pareidolia 1

Pareidolia 1 is an English-only, provenance-tracked, multi-domain dataset built for training and evaluating Firebird (a 3-class ModernBERT authorship detector with calibrated confidence).

Rather than treating AI detection as a simple binary {text, label} task, Pareidolia 1 models authorship history—preserving generator identity, source domain, editing operations, adversarial perturbations, and human–AI collaboration subtypes so detectors can be tested under model, domain, and editing shifts.

⚠️ Research & Multi-License Notice: Pareidolia 1 aggregates multiple upstream datasets carrying different terms—including CC BY-NC-SA 4.0 (e.g., EditLens, DarrenJiaImbue) and research-only / non-redistribution terms (e.g., PAN, MULTITuDE). This compilation is intended strictly for non-commercial research and benchmarking. Do not redistribute restricted upstream subsets without adhering to their original licenses.


1. Research Goal

Pareidolia investigates whether an AI text detector learns stable indicators of authorship or merely memorizes stylistic, domain, or generator artifacts:

  1. Cross-Generator Generalization: Detecting outputs from unseen model families (e.g., PaLM, Claude, Gemini, Qwen, open-weight LLMs).
  2. Editing & Mixed Authorship: Distinguishing pure human writing, pure AI writing, human-edited AI text, and AI-edited human text.
  3. Adversarial & Perturbation Robustness: Evaluating resilience against character perturbations, paraphrasing, back-translation, and refinement attacks.
  4. Domain & Style Bias: Measuring false-positive rates across news, student essays, scientific abstracts, movie reviews, creative fiction, parliamentary proceedings, and web prose.

2. Label Schema

Pareidolia 1 uses a 3-class primary label paired with a granular authorship subtype:

label Class Definition
0 Human Purely human-authored text (or pre-Nov 2022 web text marked as conservative human candidates via provenance_status).
1 AI Entirely machine-generated text.
2 Mixed Text containing both human and AI contributions (AI-edited human, human-edited AI, human-prefix/AI-continuation, or deeply interwoven text).

Granular authorship Subtypes

  • human
  • ai_generated
  • ai_edited_human
  • human_edited_ai
  • human_prefix_ai_continuation (e.g., ArguGPT GPT-2 XL)
  • mixed / deeply_mixed

3. Canonical Schema

Every record preserves normalized cross-dataset fields alongside the raw upstream row inside original_metadata:

{
  "text": "...",
  "label": 0,
  "authorship": "human",

  "source_dataset": "dataset/repo-or-id",
  "source_record_id": "original-id",
  "source_split": "train",
  "source_domain": "research_abstracts",
  "source_subdomain": "arxiv",
  "source_url": null,
  "group_id": "pair-or-document-group-id",

  "generator": "gpt-4.1",
  "generator_family": "gpt",
  "generation_date": null,
  "prompt": null,
  "instructions": null,

  "original_text": null,
  "revised_text": null,
  "editing_actor": null,
  "editing_operation": null,
  "editing_instruction": null,
  "authorship_depth": null,

  "attack_type": null,
  "attack_method": null,
  "attack_parameters": null,
  "is_attacked": false,

  "language": "en",
  "word_count": 142,
  "token_count": null,

  "original_metadata": {},
  "provenance_status": "verified"
}

4. Dataset Views & Splits

Pareidolia 1 separates training candidates, specialized analysis views, quarantined records, and sealed benchmarks:

  • main_3class — Stratified, source-capped training pool balanced across Human (0), AI (1), and Mixed (2), as well as domain, generator family, editing operation, and text length.
  • mixed_authorship — Detailed human/AI revision and collaboration records (MixSet, EditLens, BEEMO, DarrenJiaImbue, ArguGPT GPT-2 XL).
  • attacks — Perturbation, paraphrasing, and refinement records (DetectRL-X).
  • generators — Generator-attributed subset for cross-model generalization experiments.
  • authors — Author-linked records for leakage prevention and author-level evaluation.
  • domain_balanced — Domain-stratified evaluation/training view.
  • quarantine — Exact-text label conflicts, unverified pipelines, and unresolved provenance (never merged into main_3class).
  • evaluation (Sealed) — Protected official validation/test splits and benchmark-only datasets (strictly excluded from training, hard-negative mining, and calibration).

Build Summary (Auto-populated after pipeline run)

  • Total raw records inspected: [TBD]
  • Exact duplicates removed: [TBD]
  • Label-conflict hashes quarantined: [TBD]
  • Train-to-eval leakage rows removed: [TBD]
  • Final main_3class rows: [TBD] (Human: [TBD], AI: [TBD], Mixed: [TBD])

5. Source Registry

A. Protected Baseline

Source Role Notes
noumenon-labs/Earlybird-V2 Core Base Preserves cleaned baseline (MAGE, RAID, SemEval 2024 Task 8, TuringBench, HC3). validation_id and validation_ood remain strictly locked.

B. Training Candidates (Pure Human / Pure AI / Attacks)

Source Domain / Focus Handling & Caveats
sbamaker/AiHumanPolNews Political news Human news + GPT-2 generations
coai/ai-text-detection-training Research abstracts / niche models Retains model, source, domain, ArXiv metadata
heleneeriksen/gpt-vs-human-a-corpus-of-research-abstracts ArXiv abstracts Kaggle manual upload
Ateeqq/AI-and-Human-Generated-Text COVID-19 abstracts (CORD-19) Combines title + abstract
jist2009/human_ai_generated_text Student persuasive essays Preserves pair id & instructions; capped to prevent student-essay overfitting
Lyra-stellAI/AI_Human_generated_movie_reviews Movie reviews Stanford IMDB human reviews + AI counterparts
TheItCrOw/OpenLLMText Web rewrites / open LLMs Pre-2019 OpenWebText human base + paragraph rewrites
acmc/ghostbuster_essay Student essays IvyPanda human base; audited for overlap
acmc/idmgsp_abstracts_reduced Scientific abstracts IDMGSP benchmark subset
Jinyan1/COLING_2025_MGT_en Multi-domain / multi-generator Stratified by source → sub_source → model → label; deduped against Earlybird
ryuryukke/OUTFOX Paired essays & prompts Preserves problem statements + PERSUADE/LLM triplets
Rajarshi-Roy-research/Defactify_Text_Dataset Multi-model detection Retains Label_B generator attribution; official splits preserved
WUJUNCHAO/DetectRL-X Adversarial attacks English only; preserves perturbation, paraphrasing, and refinement attack metadata
zheyushen/daigt-qwen-generation Qwen essays Explicit DAIGT exception for Qwen generator coverage
kingki19/llm-generated-essay-using-palm-from-google-gen-ai Google PaLM essays Kaggle manual upload; pending human-side audit
etiennekaiser/gemini-pro-llm-daigt-dataset Gemini Pro essays Pending repo ID & DAIGT overlap audit
darraghdog/hello-claude-1000-essays-from-anthropic Claude essays Kaggle manual upload; pending schema & license check
ardavey/human-ai-generated-text Mixed / general Optional; pending Distilabel pipeline audit
SJTU-CL/ArguGPT Argumentative essays Cleaned: 6 later GPT models kept as AI; 563 GPT-2 XL essays isolated as Mixed (human_prefix_ai_continuation)
MULTITuDE Multilingual news English training split only; capped to avoid news dominance; English test split sealed (research-only)
M4 Multi-domain / multi-generator English training split only; audited against Earlybird/COLING/OUTFOX; official eval sealed
symanto/autextification2023 (optional) Multi-domain English subset only

C. Mixed-Authorship & Editing Sources

Source Categories / Contribution Handling
Dongping-Chen/MixSet Human/LLM revisions across operations English only; preserves id, category, original, revised, mixset_category, binary; official test split sealed
DarrenJiaImbue/ai-detection-demo-dataset human_written, ai_generated, ai_edited Preserves similarity scores; CC BY-NC-SA 4.0
pangram/editlens_iclr Human, AI-edited, AI-generated Training split for training; official held-out/OOD splits sealed
pangram/editlens_iclr_grammarly Grammarly-edited human text Editing-specific train/eval split separation
toloka/beemo Human, machine, expert-edited, LLM-edited Clean training partitions grouped by source answer; held-out split sealed

D. Human-Text Web Candidates

Source Scope Caveat
HuggingFaceFW/fineweb Common Crawl web prose Filtered to snapshots before Nov 2022; marked as conservative candidate, not verified authorship
Skylion007/openwebtext Pre-ChatGPT Reddit-linked web text Conservative human-web candidate

E. Zenodo & Benchmark-Only Sources

Source Domain / Task Role
Zenodo 21322627 Creative & fiction writing (~100w excerpts + AI rewrites) Official train split in training; val/test sealed
Zenodo 20611280 Parliamentary text (UK Hansard) English official train split in training; val, test & forward test sealed
Zenodo 17880411 (DET12) Detective fiction (Human vs. GPT-4.1) Evaluation-first / Benchmark
PAN 2025 (Task 2) 6 human–AI collaboration types Benchmark-only (research-only, no redistribution)
PAN 2026 Future unseen human–AI collaboration Benchmark-only
M4GT-Bench Binary, multi-way attribution, mixed-text boundary localization Benchmark-only

(Optional future additions pending live schema & overlap audit: GPT-wiki-intro, HC3+ English subset, pre-Nov 2022 IvyPanda essays, and standalone GPT-5 Real Benchmark for evaluation.)


6. Cleaning, Deduplication & Leakage Prevention

  1. English-Only Filtering: Source language metadata plus fallback language detection on multilingual corpora (MULTITuDE, M4, MixSet, DetectRL-X, Zenodo 20611280).
  2. Source-Specific Cleaning:
    • ArguGPT GPT-2 XL outputs begin with a human-written opening sentence; they are never silently truncated or labeled pure AI—they are routed to Mixed (human_prefix_ai_continuation).
    • Paired datasets (jist2009, OUTFOX, MixSet, EditLens, BEEMO) retain group_id so linked human/AI/revised texts never split across train and evaluation.
  3. Exact Deduplication & Conflict Quarantine: Texts are normalized (NFKC, whitespace-collapsed, casefolded) and hashed (SHA-256). Exact duplicates are removed using source priority (Earlybird-V2 first). Identical texts with conflicting labels are moved to quarantine.
  4. Near-Duplicate Audit: MinHash LSH flags high-similarity pairs for inspection without blindly deleting paired human/AI rewrites.
  5. Strict Benchmark Isolation: Any candidate training row matching a held-out, validation, or benchmark hash (Earlybird-V2 validation splits, PAN, M4GT-Bench, official test splits) is purged before sampling.
  6. Multi-Dimensional Stratification: Rather than sampling uniformly at random, main_3class caps dominant datasets and stratifies by authorship class, domain, generator family, source dataset, editing operation, and text length—without duplicating rare attack or editing rows.

7. Explicitly Excluded Sources

The following datasets were audited and excluded from Pareidolia 1:

  • Already covered inside Earlybird-V2: MAGE, RAID, SemEval 2024 Task 8, TuringBench, HC3.
  • Redundant / Repackaged aggregations: suraj-ranganath/unified_tell_dataset, the Pile family (artem9k, R-obi, srikanthgali), not-human-s1/ai-vs-real-text-generated-dataset, silentone0725/ai-human-text-detection-v1 (contains M4GT-Bench leakage), mehddii/ai-text-detector-v2, Yunij/testing_kaggle_ai_detection_dataset, conjuring92/ai-mix-v26, matejkore/ai-detector-dataset, and DAIGT repackagings (daigt-v2/v3/v4, Yunij/kaggle-comp-daigt, ramensoft/daigt_v3, zeyadusf, dsluciano, toshangupta, murugesann, ziq, NabeelShar, sunilthite).
  • Detector- or prediction-derived labels (not ground truth): pangram/WildAI-monthly, coai/ai-text-detection-benchmark, adaptive-classifier/ai-detector-data.
  • Synthetic human labels / poor quality / wrong task: Majkel1337/Detect-AI, Yuvrajg2107/Elite-Quality-Hybrid-Dataset, prajwaldongre/llm-detect-ai-generated-vs-student-generated-text, pratyushpuri/ai-vs-human-content-detection-1000-record-in-2025, shamimhasan8/ai-vs-human-text-dataset, bertnardomariouskono/ai-generated-text-detection-multi-model, FakeGPT, Maktabdar (no dataset).
  • Excluded by scope/availability: SFR, Telescope, AiGen-FoodReview.

8. Limitations & Known Risks

  • Unverified Human Provenance: Some retained sources (e.g., ghostbuster_essay, ardavey, FineWeb, OpenWebText) rely on upstream collection assumptions rather than cryptographic authorship proof; check provenance_status before treating subsets as gold human ground truth.
  • Domain & Register Skew: Even after capping, essay and research-abstract corpora are heavily represented relative to casual messaging or dialogue.
  • English Only: Non-English texts are excluded; models trained on Pareidolia 1 should not be evaluated on multilingual inputs.
  • Probabilistic Predictions: Detectors trained on this corpus (such as Firebird/ModernBERT) should output calibrated probabilities (e.g., via temperature scaling) and support abstention/uncertainty rather than overconfident hard labels on borderline mixed text.
Downloads last month
50