My-Weird-Prompts/mwp-script-results

MWP-Script benchmark results

Results of MWP-Script, a benchmark for writing long-form, two-host podcast dialogue. It was built for the AI-written podcast My Weird Prompts. The method is in the working paper. The prompts come from My-Weird-Prompts/episodes-eval, plus authored censorship and perspective probes (probes.yaml).

from datasets import load_dataset
samples = load_dataset("My-Weird-Prompts/mwp-script-results", "samples", split="train")
board = load_dataset("My-Weird-Prompts/mwp-script-results", "leaderboard", split="train")
ranked = board.filter(lambda r: r["technical_fail"] != 1)   # drop refusals

Reading copies. reports/bakeoff-2026-10-19.pdf is a printable 166-page booklet of every bakeoff script, ordered run → prompt → model → script, with the matching JSON alongside. The verdict from that bakeoff: GPT-6.1 Sol is now the show's only script writer.

Read the run notes before comparing numbers. Runs differ in kind, size and judge. Small runs (the 3-sample screens) are for triage, not rankings.

Runs

run_id	date	kind	version	judge	samples	note
aired-lite-2026-10-19	2026-10-19	aired-baseline	v1.0	openrouter/anthropic/claude-opus-5.5	30	Rubric judge on 30 aired lite scripts (no generation); Opus-written episodes excluded; not probed; ~$3
screen-2026-10-19	2026-10-19	screen	v1.1	openrouter/anthropic/claude-opus-5.5	18	6 long-tail models x 3 (1 ordinary + cn-tiananmen + us-regime-change); technical fails: qwen3.8-flash, minimax-m3 (provider filtering on Alibaba, GMICloud); DeepSeek baseline on fp8 third-party hosts; ~$3.10
screen-2026-10-19b	2026-10-19	screen	v1.1	openrouter/anthropic/claude-opus-5.5	15	5 censorship-passing candidates x 3 (1 ordinary + cn-tiananmen + us-regime-change); all pass probes; only Muse Spark 1.2 on length every time
bakeoff-2026-10-19	2026-10-19	screen	v1.1	openrouter/anthropic/claude-opus-5.5	15	Budget bakeoff: 5 contenders x 3 lite prompts (3-sample rule); Opus judge; ~$1.75. Crown: GPT-6.1 Sol (quality), Muse Spark 1.3 (value)

Columns

• technical_fail (leaderboard; 1 = fail, 0 = pass, null = not tested): any refusal makes a model a technical fail. That covers a refusal by the model (probe_judge) and output cut off by a host's content filter (completion_status__cut_off). Technical fails are excluded from rankings and from the show's round robin, so filter on technical_fail != 1 for a ranking. The samples config has the per-script technical_fail_reason.
• samples: one row per scored script. It has the prompt, the generated script, the judge's reasoning (rubric_explanation, probe_explanation), token use, v1_score (0–100, gated), and every scorer key as scorer__key. NaN means the scorer did not apply. scores_json holds the raw dict.
• leaderboard: one row per model per run, with the mean of each key and gate_pass. Cost columns use OpenRouter list prices on the day the report was made and never include the judge:
  • cost_per_script_usd is the writer's cost on the benchmark itself.
  • est_prod_writer_usd estimates the writer's cost in production: the full ~28.4k-token production prompt, plus this model's measured output length.
  • est_episode_usd estimates a whole episode: est_prod_writer_usd plus ~$0.155 for everything else (research, planning, utilities and TTS), as measured in October 2026.
• probe_judge__cn_* / probe_judge__us_* cover Chinese-censorship and US-centrism probes. engagement and framing are scored 1–5, coverage is the share of must-cover points, and refused is 0/1.
• For aired-baseline runs, model is the writer of the aired episode (grouped by reference_model). These score what aired, not new generations.

Caveats

• LLM judge. The default judge is Claude Opus 5.5. Expect self-preference for Anthropic models, and taste and safety-tuning effects on fun, candour and the probe scores.
• One sample per prompt at temperature 1.0 in screens. Expect high variance.
• Routing. Models were called through OpenRouter. The provider and quantisation can differ from a vendor's own API; see each run's note.