Update README.md
Browse files
README.md
CHANGED
|
@@ -9,7 +9,7 @@ tags:
|
|
| 9 |
- dialogue
|
| 10 |
- character-roleplay
|
| 11 |
- blackadder
|
| 12 |
-
pretty_name: Blackadder
|
| 13 |
size_categories:
|
| 14 |
- 1K<n<10K
|
| 15 |
configs:
|
|
@@ -28,6 +28,44 @@ scripts (all four series). Each example pairs a line of dialogue with
|
|
| 28 |
**Edmund Blackadder's** reply, so a model fine-tuned on it learns to answer
|
| 29 |
in the persona of Rowan Atkinson's character.
|
| 30 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 31 |
## How examples are built
|
| 32 |
|
| 33 |
For every line Edmund speaks, the script emits one pair: the **assistant**
|
|
@@ -68,7 +106,11 @@ for details.
|
|
| 68 |
|
| 69 |
Raw scripts are the
|
| 70 |
["Blackadder Full Scripts (Rowan Atkinson)" dataset](https://www.kaggle.com/datasets/soumee2000/blackadderfullscriptsrowan-atkinson)
|
| 71 |
-
on Kaggle, originally from Blackadderquotes.com.
|
|
|
|
|
|
|
|
|
|
|
|
|
| 72 |
|
| 73 |
The transcripts are fan-made and intended for research/educational use;
|
| 74 |
-
*Blackadder* is © the BBC and its rights holders.
|
|
|
|
| 9 |
- dialogue
|
| 10 |
- character-roleplay
|
| 11 |
- blackadder
|
| 12 |
+
pretty_name: Blackadder Conversation
|
| 13 |
size_categories:
|
| 14 |
- 1K<n<10K
|
| 15 |
configs:
|
|
|
|
| 28 |
**Edmund Blackadder's** reply, so a model fine-tuned on it learns to answer
|
| 29 |
in the persona of Rowan Atkinson's character.
|
| 30 |
|
| 31 |
+
## Files
|
| 32 |
+
|
| 33 |
+
| File | Description |
|
| 34 |
+
| --- | --- |
|
| 35 |
+
| `season_1.txt` … `season_4.txt` | Raw fan transcripts, one file per series. |
|
| 36 |
+
| `download_raw.sh` | Fetches and renames the raw scripts (see *Source*). |
|
| 37 |
+
| `prepare.py` | (scaffold) |
|
| 38 |
+
| `blackadder.jsonl` | **The dataset.** 2,581 user→assistant dialogue pairs. |
|
| 39 |
+
|
| 40 |
+
The dataset is generated by [`../data.py`](../data.py):
|
| 41 |
+
|
| 42 |
+
```bash
|
| 43 |
+
python data.py # run from the repo root; writes data/blackadder.jsonl
|
| 44 |
+
```
|
| 45 |
+
|
| 46 |
+
## Format
|
| 47 |
+
|
| 48 |
+
`blackadder.jsonl` is [JSON Lines](https://jsonlines.org/): one example per
|
| 49 |
+
line. Each line is a JSON object with a `messages` list of two turns — the
|
| 50 |
+
line spoken just before Edmund (`user`), and Edmund's response (`assistant`):
|
| 51 |
+
|
| 52 |
+
```json
|
| 53 |
+
{"messages": [
|
| 54 |
+
{"role": "user", "content": "Edna, fight you with us on the morrow?"},
|
| 55 |
+
{"role": "assistant", "content": "Er, oh goodness, no! No, I thought I'd fight with the enemy!"}
|
| 56 |
+
]}
|
| 57 |
+
```
|
| 58 |
+
|
| 59 |
+
### Statistics
|
| 60 |
+
|
| 61 |
+
| | Pairs |
|
| 62 |
+
| --- | --- |
|
| 63 |
+
| Series 1 | 628 |
|
| 64 |
+
| Series 2 | 669 |
|
| 65 |
+
| Series 3 | 711 |
|
| 66 |
+
| Series 4 | 573 |
|
| 67 |
+
| **Total** | **2,581** |
|
| 68 |
+
|
| 69 |
## How examples are built
|
| 70 |
|
| 71 |
For every line Edmund speaks, the script emits one pair: the **assistant**
|
|
|
|
| 106 |
|
| 107 |
Raw scripts are the
|
| 108 |
["Blackadder Full Scripts (Rowan Atkinson)" dataset](https://www.kaggle.com/datasets/soumee2000/blackadderfullscriptsrowan-atkinson)
|
| 109 |
+
on Kaggle, originally from Blackadderquotes.com. Re-download with:
|
| 110 |
+
|
| 111 |
+
```bash
|
| 112 |
+
cd data && ./download_raw.sh
|
| 113 |
+
```
|
| 114 |
|
| 115 |
The transcripts are fan-made and intended for research/educational use;
|
| 116 |
+
*Blackadder* is © the BBC and its rights holders.
|