---
license: mit
task_categories:
  - text-generation
  - fill-mask
tags:
  - memes
  - humor
  - byte-level
  - english
size_categories:
  - 100K<n<1M
---

# meme-text-corpus

Meme **joke text** (template names + captions + labels), structured for
template-conditioned generation of byte-level language models. Built for
training BDH (Baby Dragon Hatchling, vocab=256 raw bytes) — contains **text
only, no images**.

## Files

| File | Contents |
|---|---|
| `meme_text_corpus.txt` | Byte-level training rendering (`TEMPLATE:` / `TEXT:` / `---` records) |
| `meme_text_corpus.jsonl` | One JSON record per line with metadata (schema below) |

## Schema (JSONL)

```
{"source": "dank_learning"|"memotion2",
 "template": str,                 // meme template slug, e.g. "y u no", "brian bad news"
 "text": str,                     // caption text (single line, whitespace-normalized)
 // Memotion 2.0 rows additionally carry:
 "humor"?: str, "sarcasm"?: str, "offensive"?: str, "sentiment"?: str, "split"?: str}
```

Text rendering format (`.txt`):

```
TEMPLATE: y u no
TEXT: steve jobs y u no respawn?!
---
```

## Sources & licenses

| Source | License | Contribution |
|---|---|---|
| [Dank Learning](https://github.com/alpv95/Dank-Learning) (memegenerator.net scrape, 2018; paper arXiv:1806.04510) | MIT | ~411k template-caption pairs (majority of corpus) |
| [Memotion 2.0](https://huggingface.co/datasets/Ahren09/MMSoc_Memotion) | Research dataset — see its dataset card; labels used as metadata only | ~7k OCR'd meme texts with humor/sarcasm/offensive/sentiment labels |
| imgflip `get_memes` API | imgflip ToS | 100-template taxonomy reference (metadata only, no caption text) |

## Cleaning

- Exact dedup on whitespace/case-normalized text; near-dedup with word-3-gram
  Jaccard ≥ 0.8 within (template, first-3-words) buckets.
- Filters (excluded, counts in `CORPUS_REPORT.md`): non-English heuristic,
  slur/hateful blocklist, explicit-sexual-content list, Memotion
  `very_offensive` rows, length outside [8, 600] chars.
- Mild profanity remains — meme text is crude by nature. The `offensive` label
  is preserved for downstream filtering.

## Known limitations

- 2018-era meme culture (memegenerator.net); dated references.
- Captions are single-line (top/bottom split not recoverable from the source).
- English-only by construction; non-English filter is heuristic.

## Reproduction

Full pipeline + scripts: see the `create-datasets` project repo
(`scrapers/`, `scripts/`, `qc/`, `CORPUS_REPORT.md`).