---
license: other
language:
- en
tags:
- perplexity
- evaluation
- news
- time-isolated
configs:
- config_name: default
  data_files:
  - split: "00"
    path: "data/00.parquet"
  - split: "01"
    path: "data/01.parquet"
  - split: "02"
    path: "data/02.parquet"
  - split: "03"
    path: "data/03.parquet"
  - split: "04"
    path: "data/04.parquet"
  - split: "05"
    path: "data/05.parquet"
  - split: "06"
    path: "data/06.parquet"
  - split: "07"
    path: "data/07.parquet"
  - split: "08"
    path: "data/08.parquet"
  - split: "09"
    path: "data/09.parquet"
  - split: "10"
    path: "data/10.parquet"
  - split: "11"
    path: "data/11.parquet"
  - split: "12"
    path: "data/12.parquet"
  - split: "13"
    path: "data/13.parquet"
  - split: "14"
    path: "data/14.parquet"
  - split: "15"
    path: "data/15.parquet"
  - split: "16"
    path: "data/16.parquet"
  - split: "17"
    path: "data/17.parquet"
  - split: "18"
    path: "data/18.parquet"
  - split: "19"
    path: "data/19.parquet"
  - split: "20"
    path: "data/20.parquet"
  - split: "21"
    path: "data/21.parquet"
  - split: "22"
    path: "data/22.parquet"
  - split: "23"
    path: "data/23.parquet"
  - split: "24"
    path: "data/24.parquet"
  - split: "25"
    path: "data/25.parquet"
---

# TheFinAI/ppl — Per-Year News Corpus for Perplexity Evaluation

A held-out news corpus for evaluating per-year perplexity of time-isolated
language models (TILE / MoE per-year experts).

## Structure

- One split per year, named `00` (= 2000) through `25` (= 2025).
- Each split contains ~2000 news articles sampled month-stratified
  (~167/month) from The Guardian Open Platform.
- Sections kept: world, us-news, uk-news, business, politics, technology,
  science, environment, society, education, media, law, etc.
  (Sport, lifestyle, food, fashion, film, books, travel, games dropped.)
- All articles passed `token_count >= 50` (cl100k_base) filter.

## Schema

| column        | dtype  | note                                  |
|---------------|--------|---------------------------------------|
| `source`      | str    | constant `"guardian"`                 |
| `date`        | int32  | year, e.g. `2003`                     |
| `pub_date`    | str    | publication date `YYYY-MM-DD`         |
| `section`     | str    | guardian sectionId                    |
| `headline`    | str    | webTitle                              |
| `url`         | str    | webUrl                                |
| `text`        | str    | full article body (`bodyText`)        |
| `token_count` | int32  | tiktoken `cl100k_base` token count    |

## Usage

```python
from datasets import load_dataset
ds = load_dataset("TheFinAI/ppl", split="03")  # year 2003
```

## Notes

- Cross-year fairness: all years drawn from the same source pipeline, same
  section filter, same length filter, same tokenizer. PPL comparisons
  across years should not be biased by corpus shift.
- Single-source: only Guardian. For multi-outlet eval, supplement with NYT
  metadata (`TheFinAI/ppl_nyt` if/when added).