dadbench / README.md
akgrenier's picture
Add code + sibling model links
c1b3fd1 verified
|
Raw History Blame Contribute Delete
3.1 kB
---
license: apache-2.0
language:
- en
pretty_name: DadBench
size_categories:
- n<1K
task_categories:
- text-generation
tags:
- humor
- evaluation
- benchmark
- llm-as-a-judge
- reward-modeling
- rlhf
configs:
- config_name: default
data_files:
- split: human_ratings
path: data/human_ratings.jsonl
---
# DadBench
**A benchmark for dad-joke quality across language models — and a case study in how hard it is to measure a subjective objective without getting Goodharted.**
DadBench scores model-written dad jokes on one question: *would a real dad be proud to inflict this?* It exists because the [DadBench project](https://aimatey.co/dadbench/) RL-trained a model (**PunTune**, a fine-tune of Thinking Machines' Inkling) to tell dad jokes, and needed an evaluation rigorous enough to tell whether training actually worked. The answer — *training didn't teach new wit, it taught the model to pick a punchline and stop* — only held up after the eval itself survived four instrument rebuilds and several fired judges.
📝 **Full write-up + live arena:** https://aimatey.co/dadbench/
## What's in here
| File | Contents |
|---|---|
| `data/human_ratings.jsonl` | **143 single-rater human ratings** (the calibration ground truth). Loads as the `human_ratings` split. |
| `leaderboard.json` | Final model board — overall + per-track scores in permanent ladder units. |
| `ladder.json` | The **frozen anchor ladder**: per-topic rungs a new model's jokes are placed against, so scores never drift. |
| `joke_gallery.md` | Top and bottom jokes, verbatim, with scores — the qualitative texture. |
### `human_ratings.jsonl` fields
`sitting` (rating session id) · `n` (index) · `joke` (verbatim text) · `source` (which model/round produced it) · `topic` · `rating` (1–5, blind).
## Method in one paragraph
Jokes are scored across four tracks (plain topics, cultural references, comparisons, and long-context news articles). Each joke is placed in **9 blind, order-randomized rankings** against a **frozen ladder** of anchor jokes, and strength is a Bradley–Terry fit — so adding a model or checkpoint never moves anyone else's number. The judge is a **reasoning-parity panel** selected by correlation against the human ratings in this dataset. The project's recurring finding is that the reward is a Goodhart machine: any criterion every candidate satisfies goes invisible under group-relative RL, and uncontrolled thinking-token budgets silently reorder the board. That makes DadBench useful less as a "who's funniest" leaderboard and more as a **testbed for judge robustness and reward design**.
## Related
- **Model:** [`aimatey/PunTune-0.6`](https://huggingface.co/aimatey/PunTune-0.6) · [`aimatey/PunTune-0.5`](https://huggingface.co/aimatey/PunTune-0.5), the Inkling fine-tunes this benchmark was built to measure.
- **Code:** https://github.com/aimatey-co/dadbench
- **Write-up + interactive arena:** https://aimatey.co/dadbench/
## License
Apache-2.0. Built on [Thinking Machines' Inkling](https://huggingface.co/thinkingmachines/Inkling) (also Apache-2.0).