yoonholee/style-eval-pg-karpathy-gwern
Style Eval Corpus Writing from 11 internet writers with instantly recognizable but distinct styles. Built for style-as-reward-inference experiments: given a writer's body of work, infer their implicit reward function and encode it as eval components (rubrics, classifiers, probes). Contents Author Pieces Words Register Source Paul Graham 229 562,990 Contrarian startup essays paulgraham.com Andrej Karpathy 34 101,337 Tutorial-as-thinking-aloud… See the full description on the dataset page: https://huggingface.co/datasets/yoonholee/style-eval-pg-karpathy-gwern.
Style Eval Corpus
Writing from 11 internet writers with instantly recognizable but distinct styles. Built for style-as-reward-inference experiments: given a writer's body of work, infer their implicit reward function and encode it as eval components (rubrics, classifiers, probes).
Contents
Design
The corpus spans several deliberate contrasts:
- Long-form analytical (PG, Karpathy, Gwern, Scott Alexander, Ceglowski, Joel, Eliezer) vs short-form (dril, Trump, Sivers, Naval). Same eval framework, different registers.
- Same genre, different rewards: PG / Karpathy / Gwern / Scott Alexander / Eliezer all write tech/rationalist essays but with radically different implicit reward functions. Within-genre discrimination is the hard task.
- Corpus size varies deliberately: Karpathy (34 posts) and Naval (52 chapters) are the few-shot setting. Trump (46K tweets) and Eliezer (1.5K posts) are data-rich. Tests generalization from thin vs thick reference sets.
Schema
License
- Gwern Branwen: CC-0 (public domain).
- Naval Ravikant: Almanack is CC-licensed, explicitly free.
- Eliezer Yudkowsky: LessWrong content is CC BY 4.0.
- dril / Trump tweets: sourced from existing HF datasets.
- All others: publicly posted writing, research use under fair use.
Curation/parsing released under CC0.