---
license: cc-by-nc-4.0
language:
  - en
task_categories:
  - text-generation
  - fill-mask
  - feature-extraction
task_ids:
  - language-modeling
tags:
  - ao3
  - fanfiction
  - creative-writing
  - long-form
  - literature
  - english
  - pretraining
  - chatml
  - openhermes
  - alignment
pretty_name: AO3-2020 Fanfiction Corpus
size_categories:
  - 1B<n<10B
source_datasets:
  - original
multilinguality: monolingual
dataset_size_bytes: 539375493120
configs:
  - name: full
num_examples: 1553596

• name: tokens_2_5M num_examples: 711

• name: tokens_1_25M num_examples: 392

• name: tokens_250K num_examples: 76

• name: tokens_100K num_examples: 25

• name: tokens_75K num_examples: 20

• name: tokens_50K num_examples: 14

• name: tokens_37_5K num_examples: 10

• name: tokens_25K num_examples: 6

• name: tokens_10K num_examples: 2

• name: tokens_7_5K num_examples: 2

• name: tokens_5K num_examples: 1

• name: tokens_2_5K num_examples: 0

• name: three_quarters num_examples: 1165197

• name: seven_tenths num_examples: 1087516

• name: five_ninths num_examples: 863108

• name: half num_examples: 776798

• name: quarter num_examples: 388399

• name: fifth num_examples: 310719

• name: eighth num_examples: 194199

• name: twelfth num_examples: 129466

• name: fifteenth num_examples: 103573

• name: sixteenth num_examples: 97100

• name: twenty_fourth num_examples: 64733

• name: aug_openhermes_10K num_examples: 2

• name: aug_openhermes_37_5K num_examples: 10

• name: aug_openhermes_100K num_examples: 25

• name: aug_openhermes_250K num_examples: 76 dataset_info: features: - name: text dtype: string splits: - name: train num_examples: 15993173 annotations_creators: [] language_creators: - crowdsourced

  AO3-2020 — Archive of Our Own Fanfiction Corpus

  ---

  📋 Overview

  Field	Value
  Total documents	15,993,173
  Total characters	~21,816,825,882
  Estimated tokens	~5,454,206,470
  Raw source size	502 GB (SQLite)
  Language	English
  Source	Archive of Our Own (ao3.org)
  Cutoff date	2020
  License	CC BY-NC 4.0
  Parquet compression	ZSTD level 3
  Last updated	2026-05-17

  ---

  🚀 Quick start

  from datasets import load_dataset
  
  # Full 502 GB corpus — use streaming=True for large machines
  ds = load_dataset("ray0rf1re/AO3-2020", "full", streaming=True)
  
  # Fast experiment — 75 K token slice
  ds = load_dataset("ray0rf1re/AO3-2020", "tokens_75K")
  
  # SFT-ready ChatML format (OpenHermes 2.5 template)
  ds = load_dataset("ray0rf1re/AO3-2020", "aug_openhermes_100K")
  
  # Deterministic 1/8 uniform sample
  ds = load_dataset("ray0rf1re/AO3-2020", "eighth")
  
  # Print first document
  for ex in ds["train"]:
      print(ex["text"][:500])
      break
  

  ---

  🗂️ Subsets

  Full corpus

  Subset	Rows
  full	1,553,596

  Token-capped subsets

  Rows are taken in source order until the cumulative estimated token budget (≈ 4 chars / token) is reached. Good for controlled-scale experiments.

  Subset	Token cap	Rows
  tokens_2_5M	2,500,000	711

| tokens_1_25M | 1,250,000 | 392 | | tokens_250K | 250,000 | 76 | | tokens_100K | 100,000 | 25 | | tokens_75K | 75,000 | 20 | | tokens_50K | 50,000 | 14 | | tokens_37_5K | 37,500 | 10 | | tokens_25K | 25,000 | 6 | | tokens_10K | 10,000 | 2 | | tokens_7_5K | 7,500 | 2 | | tokens_5K | 5,000 | 1 | | tokens_2_5K | 2,500 | 0 |

### Fraction subsets

Deterministic, uniform row-skip sampling — every `d`th row starting at 0.
Reproducible across library versions without a fixed seed.

| Subset | Fraction | Coverage | Rows |
|--------|----------|----------|------|
| `three_quarters` | 3/4 | 75.00 % | 1,165,197 |

| seven_tenths | 7/10 | 70.00 % | 1,087,516 | | five_ninths | 5/9 | 55.56 % | 863,108 | | half | 1/2 | 50.00 % | 776,798 | | quarter | 1/4 | 25.00 % | 388,399 | | fifth | 1/5 | 20.00 % | 310,719 | | eighth | 1/8 | 12.50 % | 194,199 | | twelfth | 1/12 | 8.33 % | 129,466 | | fifteenth | 1/15 | 6.67 % | 103,573 | | sixteenth | 1/16 | 6.25 % | 97,100 | | twenty_fourth | 1/24 | 4.17 % | 64,733 |

### Augmented subsets — OpenHermes 2.5 ChatML template

Same token budgets as their plain counterparts, but each document is
wrapped in the **OpenHermes 2.5 / ChatML** instruction template so the
data can be fed directly to SFT trainers (e.g. LLaMA-Factory, Axolotl,
Unsloth) without additional preprocessing.

Template applied to every row:
```
<|im_start|>system
You are a creative writing assistant specialising in long-form fanfiction. You produce engaging, vivid, and high-quality stories.<|im_end|>
<|im_start|>user
Write a fanfiction story.<|im_end|>
<|im_start|>assistant
{text}<|im_end|>
```

| Subset | Token cap | Rows |
|--------|-----------|------|
| `aug_openhermes_10K` | 10,000 | 2 |

| aug_openhermes_37_5K | 37,500 | 10 | | aug_openhermes_100K | 100,000 | 25 | | aug_openhermes_250K | 250,000 | 76 |

---

## 📁 Data schema

```json
{
    "text":  "string — UTF-8 text of the fanfiction work or chapter.",
    "...":   "All additional metadata columns from the source SQLite are preserved as-is."
}
```

Augmented subsets replace `text` with the ChatML-formatted version.

---



The corpus is directly readable with:
- 🤗 `datasets` (streaming or map-style)
- DuckDB: `SELECT * FROM read_parquet('data/full/*.parquet')`
- Polars: `pl.scan_parquet('data/full/*.parquet')`
- Spark: `spark.read.parquet('data/full/')`
- PyArrow: `pq.read_table('data/full/')`

---

## ⚖️ License & ethics

- **License**: [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)
- Non-commercial research use only.
- Authors retain copyright over their individual works.
- Works rated **Explicit** or **Mature** may be present in the `full` corpus.
- If you are an AO3 author and wish your work removed, please open an issue.

---

## 🔗 Citation

```bibtex
@dataset{ao3_2020,
  author    = {ray0rf1re},
  title     = {AO3-2020: Archive of Our Own Fanfiction Corpus},
  year      = {2020},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/ray0rf1re/AO3-2020}
}
```

---

⚠️ adult_18plus_100K — SAFETY / ALIGNMENT SUBSET

❌ DO NOT train a generative model on this subset as ordinary SFT data.

✅ Use it inverted — as the rejected side of preference pairs.

This subset contains the first ~100,000 tokens of content rated Explicit or Mature on AO3 (26 chapters).

Technique	How to use this subset
DPO / RLHF	Label completions as rejected; safe rewrites as chosen.
Content-safety classifier	Negative class in binary classifier training.
Inverted SFT	Generate from this data, teach the model to rewrite it safely.

Training on this data naively produces an unsafe model. Training against it — using it as the "what not to do" signal — produces a safer model than one that never saw this kind of content.