README.md

---
language:
- en
license: other
license_name: public-domain-gutenberg-texts-and-cc0-metadata
license_link: LICENSE
task_categories:
- text-generation
pretty_name: Talkie YaRN 32k Gutenberg Pre-1931 265M
size_categories:
- 1K<n<10K
tags:
- talkie
- project-gutenberg
- pre-1931
- long-context
- yarn
configs:
- config_name: default
  data_files:
  - split: train
    path: train.jsonl
  - split: validation
    path: validation.jsonl
---

# Talkie YaRN 32k Gutenberg Pre-1931 265M

This dataset is the continued-pretraining corpus used for the Talkie YaRN 32k
context-extension experiments, including the recommended checkpoint
[`xlr8harder/talkie-1930-13b-yarn-32k-tf`](https://huggingface.co/xlr8harder/talkie-1930-13b-yarn-32k-tf).

It was generated from
[`common-pile/project_gutenberg_filtered`](https://huggingface.co/datasets/common-pile/project_gutenberg_filtered),
joined with a Gutenberg publication-year metadata table from Kaggle,
[`yuvalschwartz/gutenberg-book-metadata-with-publication-years`](https://www.kaggle.com/datasets/yuvalschwartz/gutenberg-book-metadata-with-publication-years),
then filtered to English public-domain books with publication years from 1500
through 1930. Project Gutenberg boilerplate markers were stripped and each book
was capped at 500,000 characters.

## Splits

| Split | Rows | Talkie tokens | Notes |
| --- | ---: | ---: | --- |
| train | 3,146 | 265,080,702 | Used for continued pretraining |
| validation | 65 | 5,074,819 | Held-out continuation after the train manifest |

The train JSONL was reconstructed from the surviving manifest after the original
local file was lost. The reconstruction preserved the original Gutenberg ID
order and per-book metadata from `train.summary.json`.

## Files

- `train.jsonl`: training split, one book per line.
- `validation.jsonl`: validation split, one book per line.
- `train.summary.json`: train manifest with per-book metadata and Talkie token counts.
- `validation.summary.json`: validation manifest with per-book metadata and Talkie token counts.
- `scripts/prepare_gutenberg_with_years.py`: generation and reconstruction script.
- `metadata/gutenberg_publication_years.csv`: publication-year metadata used by the script.
- `logs/prepare_train.log`: reconstruction log for the train split.

Each JSONL row has this shape:

```json
{"text": "...", "meta": {"gutenberg_id": "...", "publication_year": 1900, "...": "..."}}
```

## Generation Command

The train split was reconstructed with:

```bash
python scripts/prepare_gutenberg_with_years.py \
  --dataset common-pile/project_gutenberg_filtered \
  --split train \
  --metadata-csv metadata/gutenberg_publication_years.csv \
  --out train.jsonl \
  --summary-out train.summary.json \
  --include-summary train.summary.json \
  --reuse-include-summary-token-counts \
  --min-year 1500 \
  --max-year 1930 \
  --languages en \
  --rights-contains "public domain" \
  --min-chars 100000 \
  --max-chars-per-book 500000 \
  --target-talkie-tokens 265000000 \
  --max-books 0
```

For a fresh build rather than manifest reconstruction, omit `--include-summary`
and `--reuse-include-summary-token-counts`, and write to a new output path.

The validation split used the same filters while excluding the train manifest:

```bash
python scripts/prepare_gutenberg_with_years.py \
  --dataset common-pile/project_gutenberg_filtered \
  --split train \
  --metadata-csv metadata/gutenberg_publication_years.csv \
  --out validation.jsonl \
  --summary-out validation.summary.json \
  --exclude-summary train.summary.json \
  --min-year 1500 \
  --max-year 1930 \
  --languages en \
  --rights-contains "public domain" \
  --min-chars 100000 \
  --max-chars-per-book 500000 \
  --target-talkie-tokens 5000000 \
  --max-books 0
```

## Integrity

- `train.jsonl` SHA-256: `50a0908c33e741ce96e9b8c75f266369aac7f1ccd76757970904031b487c283c`
- `validation.jsonl` SHA-256: `efa8f3218ec219aeb5ffd5cbe70b92c3bb42e4b599dbb74f3dde7dc4b256d21a`

## License Notes

This repository uses `license: other` with a custom license description because
it combines multiple upstream sources. The text rows inherit the public-domain
status reported by `common-pile/project_gutenberg_filtered` for the selected
Project Gutenberg rows. The included publication-year metadata CSV comes from
the Kaggle dataset
[`yuvalschwartz/gutenberg-book-metadata-with-publication-years`](https://www.kaggle.com/datasets/yuvalschwartz/gutenberg-book-metadata-with-publication-years),
which is listed as CC0: Public Domain. See [`LICENSE`](./LICENSE) for the
component-level license notes.