---
license: cc-by-nc-4.0
language:
- en
task_categories:
- text-generation
tags:
- conversational
- dialogue
- character-roleplay
- blackadder
pretty_name: Blackadder Conversation
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: blackadder.jsonl
---

# Blackadder Chat Dataset

<img src="https://i.pinimg.com/1200x/5a/22/ab/5a22ab0bd42e793ee569a73a3b1e7aaf.jpg" alt="Blackadder" width="400">

A chat-style instruction dataset distilled from the full *Blackadder* TV
scripts (all four series). Each example pairs a line of dialogue with
**Edmund Blackadder's** reply, so a model fine-tuned on it learns to answer
in the persona of Rowan Atkinson's character.

## Files

| File | Description |
| --- | --- |
| `season_1.txt` … `season_4.txt` | Raw fan transcripts, one file per series. |
| `download_raw.sh` | Fetches and renames the raw scripts (see *Source*). |
| `prepare.py` | (scaffold) |
| `blackadder.jsonl` | **The dataset.** 2,581 user→assistant dialogue pairs. |

The dataset is generated by [`../data.py`](../data.py):

```bash
python data.py    # run from the repo root; writes data/blackadder.jsonl
```

## Format

`blackadder.jsonl` is [JSON Lines](https://jsonlines.org/): one example per
line. Each line is a JSON object with a `messages` list of two turns — the
line spoken just before Edmund (`user`), and Edmund's response (`assistant`):

```json
{"messages": [
  {"role": "user", "content": "Edna, fight you with us on the morrow?"},
  {"role": "assistant", "content": "Er, oh goodness, no! No, I thought I'd fight with the enemy!"}
]}
```

### Statistics

| | Pairs |
| --- | --- |
| Series 1 | 628 |
| Series 2 | 669 |
| Series 3 | 711 |
| Series 4 | 573 |
| **Total** | **2,581** |

## How examples are built

For every line Edmund speaks, the script emits one pair: the **assistant**
turn is Edmund's line, and the **user** turn is the line spoken immediately
before it by *another* character.

Key decisions made while parsing the (messy, hand-typed) transcripts:

- **Edmund is matched by all of his aliases, resolved per series.** The
  scripts abbreviate speakers differently in each episode, and the
  abbreviations *conflict* — e.g. `B` means **Baldrick** in series 2 & 3, not
  Blackadder. The alias sets used are:

  | Series | Edmund labels |
  | --- | --- |
  | 1 | `Edmund`, `EBA`, `BA`, `E` |
  | 2 | `Edmund`, `Blackadder`, `BA`, `E` |
  | 3 | `Edmund`, `Blackadder`, `E` |
  | 4 | `Edmund`, `Blackadder` |

- **The user turn is the nearest preceding *other*-character line.**
  Intervening stage directions are skipped, and consecutive Edmund lines are
  merged into a single assistant turn.
- **Stage directions are stripped** — `[...]`, `(...)`, and `{...}` (one
  episode uses braces) — including directions that wrap across script lines.
- **Episode boundaries** (`Sharing is caring!` + intro paragraph) and the
  per-episode **cast legends** (`B: Baldrick`) are detected and excluded.
- **Wrapped dialogue and scene narration** are not mistaken for new speaker
  turns; a line is only treated as a new speaker when its label is short and
  capitalised.

The parser is heuristic, so a small amount of residual noise remains (e.g. a
rare line where two characters share one transcript line keeps an inline
`Name:` tag). See the docstring and comments in [`../data.py`](../data.py)
for details.

## Source

Raw scripts are the
["Blackadder Full Scripts (Rowan Atkinson)" dataset](https://www.kaggle.com/datasets/soumee2000/blackadderfullscriptsrowan-atkinson)
on Kaggle, originally from Blackadderquotes.com. Re-download with:

```bash
cd data && ./download_raw.sh
```

The transcripts are fan-made and intended for research/educational use;
*Blackadder* is © the BBC and its rights holders.