My-Weird-Prompts/Model-Chronicle

My Weird Prompts — Model Chronicle

My Weird Prompts is an AI-generated podcast: every episode's dialogue is written by a large language model, and the model that writes it has changed many times since November 2025. The back catalogue is therefore a chronicle of how frontier models evolved, with each era's writer on air. This dataset is a small, curated slice of it — up to three aired scripts per writer model — so anyone can read eras side by side ("Gemini 2.5 Flash vs DeepSeek V4 Pro vs GPT-6.1 Sol") without downloading the full corpus.

59 scripts from 28 writer models, built 2026-10-24.

• Site hub, with the leaderboard and writer history: https://myweirdprompts.com/benchmark
• Full corpus (every episode, turns, text statistics): My-Weird-Prompts/transcripts
• Browse and filter every transcript by model: My-Weird-Prompts/transcripts-explorer
• Controlled, same-prompt benchmark of candidate writers: https://github.com/My-Weird-Prompts/eval

Read this before comparing

These are not same-prompt comparisons. Each script answers whatever listener or host prompt was on the queue that day; topics, lengths and episode formats differ. For like-for-like scoring of writers on identical briefs, use the eval repo above.

The model is not the only thing that changed between eras:

• The system-prompt template changed with the model, twice. In the full catalogue (measured 2026-09-22 over 5,332 episodes), template 3.5 → 4.0 landed on the April 2026 Gemini → DeepSeek switch, and 4.0 → 4.1 on the September 2026 DeepSeek V4 Pro → V4.1 Flash switch. A difference between those eras is never attributable to the model alone. prompt_version is included per row; null means not recorded (before March 2026), not "none".
• A deterministic verbal-tic cleanup pass runs after generation and deletes stock phrasings (conjunction openers, sycophantic openers, echo handoffs), so the aired text is not the raw model output. The full corpus counts what it removed per episode (tic_* columns).
• Episode length moved: median length rose from about 3,000 words (2025) to about 4,200 (2026), so length-sensitive measures (n-gram repetition, compression) trend partly with length.
• Some eras are a few days long. The March 2026 and October 2026 bake-offs aired one to a handful of episodes per candidate; those share that week's prompts, so anything topic-linked in them is subject matter, not style. model_episode_count shows how much each model actually wrote.
• Offline scripts (written = offline, 6 rows here) were written in Claude Code and submitted as finished scripts rather than by the production pipeline. They are used only for models with no pipeline-written episode.

Selection rule

Per canonical writer model, from aired English episodes with a transcript (test episodes, translations and audio_eval specials excluded; records that name no model — manual-script, manual-generation, unknown, none, openrouter/mixed — skipped):

1. Pool = the first non-empty of: pipeline-written with at least 1,500 words; pipeline-written of any length; offline-written with at least 1,500 words; offline-written of any length (the pool column says which).
2. Picks, by publication date within the pool: the earliest, the median-date one and the latest. A model with one or two eligible episodes gets one or two.

The build is deterministic. It is rebuilt on demand by scripts/build_model_chronicle.py (a copy of the exact code ships in this repo) from the production database; hand edits on the Hub are overwritten.

Models

Canonical id	Name	Provider	Era (first → last aired)	Episodes	Picks	Pool
gemini-2-5-flash	Gemini 2.5 Flash	google	2025-11-24 → 2026-03-20	66	3	pipeline
gemini-3-flash	Gemini 3 Flash	google	2025-12-21 → 2025-12-23	27	3	pipeline
gemini-3-flash-preview	Gemini 3 Flash Preview	google	2025-12-23 → 2026-04-11	1,941	3	pipeline
claude-sonnet-4-6	Claude Sonnet 4.6	anthropic	2026-03-26 → 2026-04-21	202	3	pipeline
glm-5-turbo	GLM-5 Turbo	z-ai	2026-03-26 → 2026-03-26	1	1	pipeline
mimo-v2-omni	Xiaomi MiMo v2 Omni	xiaomi	2026-03-26 → 2026-03-26	1	1	pipeline
mimo-v2-pro	Xiaomi MiMo v2 Pro	xiaomi	2026-03-26 → 2026-03-29	4	3	pipeline
minimax-m2-7	MiniMax M2.7	minimax	2026-03-26 → 2026-03-29	3	3	pipeline
deepseek-v3-2	DeepSeek v3.2	deepseek	2026-03-28 → 2026-04-24	38	3	pipeline
glm-5	GLM-5	z-ai	2026-03-28 → 2026-03-28	1	1	pipeline
grok-4-1-fast	Grok 4.1 Fast	x-ai	2026-03-28 → 2026-03-28	1	1	pipeline
mercury-2	Mercury 2	inception	2026-03-28 → 2026-03-28	1	1	pipeline
mimo-v2-flash	Xiaomi MiMo v2 Flash	xiaomi	2026-03-28 → 2026-03-28	1	1	pipeline
deepseek-chat	DeepSeek v3.2 (deepseek-chat)	deepseek	2026-04-16 → 2026-04-17	19	3	pipeline
deepseek-v4-pro	DeepSeek v4 Pro	deepseek	2026-04-25 → 2026-09-20	2,957	3	pipeline
deepseek-flash	DeepSeek 4.1 Flash	deepseek	2026-09-20 → 2026-10-19	653	3	pipeline
claude-haiku-5-5	Claude Haiku 5.5	anthropic	2026-10-17 → 2026-10-21	10	1	pipeline
claude-opus-5-5	Claude Opus 5.5	anthropic	2026-10-17 → 2026-10-19	43	3	offline
glm-5-2	GLM-5.2	z-ai	2026-10-17 → 2026-10-19	3	3	pipeline
gpt-6-luna	GPT-6 Luna	openai	2026-10-17 → 2026-10-17	1	1	pipeline
ling-3-1-flash	Ling 3.1 Flash	inclusionai	2026-10-17 → 2026-10-18	2	2	pipeline
mimo-v2-6-pro	Xiaomi MiMo V2.6 Pro	xiaomi	2026-10-17 → 2026-10-17	1	1	pipeline
solar-pro4	Solar Pro 4	upstage	2026-10-17 → 2026-10-17	1	1	pipeline
claude-sonnet-5-5	Claude Sonnet 5.5	anthropic	2026-10-19 → 2026-10-23	77	3	offline
glm-5-3	GLM-5.3	z-ai	2026-10-19 → 2026-10-19	1	1	pipeline
glm-5-3-flash	GLM-5.3 Flash	z-ai	2026-10-19 → 2026-10-19	1	1	pipeline
gpt-6-1-sol	GPT-6.1 Sol	openai	2026-10-19 → 2026-10-24	18	3	pipeline
muse-spark-1-3	Muse Spark 1.3	meta	2026-10-19 → 2026-10-23	3	3	pipeline

Columns

Column	Description
model_canonical	Canonical writer id: vendor prefix dropped, . folded to - (same rule as llm_model_canonical in My-Weird-Prompts/transcripts)
llm_model	The writer id as recorded on the episode
provider	Vendor, from the recorded prefix or the bare name
display_name	Human label used on myweirdprompts.com
era_first	Date of the model's first aired episode (all of its episodes, not just this pool)
era_last	Date of the model's latest aired episode
model_episode_count	Aired episodes the model is credited with
pick	earliest, middle (median date within the pool) or latest
pool	Which pool the pick came from: pipeline, pipeline-short, offline or offline-short (short = under 1,500 words)
episode_number	Display number at build time; not a stable id (use slug)
slug	Stable episode id
title	Episode title
pub_date	Publication date, YYYY-MM-DD (UTC)
written	pipeline (the production pipeline wrote it) or offline (written in Claude Code and submitted as a finished script)
episode_type	Pipeline episode type
prompt_version	System-prompt template version, where recorded (null before March 2026: not recorded, not none)
pipeline_version	Pipeline version string, where recorded
word_count	Whitespace-delimited words in the transcript, speaker labels included
url	Episode page
transcript_url	Plain-text transcript on the site
transcript	The aired script, Speaker: text per turn

Canonicalisation: anthropic/claude-sonnet-4.6 and claude-sonnet-4-6 are the same model under two spellings and share model_canonical = claude-sonnet-4-6. deepseek-chat is kept as its own id here, as in the full corpus; those episodes say on air that DeepSeek v3.2 wrote them.

Use

from datasets import load_dataset
ds = load_dataset("My-Weird-Prompts/model-chronicle", split="train")

The writing models' providers set their own terms on using model output, including for training other models; check them before fine-tuning on these scripts. For a fine-tuning set curated with that in mind, contact show@myweirdprompts.com.

License

CC-BY-4.0. Attribute to My Weird Prompts / Daniel Rosehill.