The whole show, as open data
Two datasets on Hugging Face, both CC-BY-4.0, both rebuilt daily from the production database. Between them: every episode we have published, every transcript, and the machinery that made them.
The production record
One row per episode: metadata, the full transcript, links back to the published episode, and the generation telemetry — which model wrote the script, pipeline version, GPU, timings, compute cost.
The textual-analysis corpus
The same transcripts, shaped for linguistics: narrowed metadata, per-episode text statistics, repetition measures, and a second config splitting the dialogue into speaker turns.
data-dictionary.json
describes every field in both datasets — type, meaning, and the share of rows
actually populated, measured from the published data rather than asserted.
Where a null carries meaning, it says what the null means.
Check coverage before you use a column. Several fields are well under 50 % populated, and the gaps are not random — they follow eras of the show. A field that is empty for the back catalogue is missing data, not a measured zero.
The dataset navigator
is the same file as a browsable page — search and filter every field, see
coverage at a glance, and link someone to a single column with a
#field= anchor.
22.6 million words of LLM-written conversation by 14 recurring speakers, with turn-level segmentation. A corpus of machine dialogue at a scale and consistency that is hard to assemble from scraped sources — one show, one format, one cast, 5,300 episodes.
Every turn is timestamped and attributed to a script-writing model. You can follow a phrase — "here’s the thing", "at the end of the day" — as a rate per 10,000 words across eras, and see where in an episode it tends to land.
Per-episode rep-4 and rep-8, gzip compression ratio, repeated-span coverage and adjacent-turn similarity, each localised to the first, middle and last third. Enough to ask whether a model pads when it runs out of material — the target length it was writing toward is in the data too.
Flesch scores, MATTR, hapax ratio and words-per-minute against actual audio duration, so written complexity can be compared with delivered pace.
Per-episode GPU type, worker count, synthesis time and compute cost in USD. A rare public record of what it actually costs to generate and voice long-form audio at scale.
Prompt-free, permissively licensed, consistently formatted long-form dialogue. Useful as a conversational-style reference set, or as evaluation data for summarisation and speaker-attribution tasks.
from datasets import load_dataset
episodes = load_dataset("My-Weird-Prompts/transcripts", "episodes", split="train")
turns = load_dataset("My-Weird-Prompts/transcripts", "turns", split="train")
# pub_date is ISO 8601 at mixed precision — pass format="ISO8601" or pandas
# will infer one format from the first row and coerce most of the column to NaT
import pandas as pd
df = episodes.to_pandas()
df["pub_date"] = pd.to_datetime(df["pub_date"], utc=True, format="ISO8601") main is rebuilt daily, so a figure measured against it cannot be
reproduced tomorrow. For a number that stays put, use a dated snapshot under
exports/
or pin a commit revision.
Figures on this page are as of 22 September 2026; both datasets grow daily.