Stella Biderman

Open-source work

Biderman leads EleutherAI, whose staff page describes researchers, engineers, and operators supporting open AI science across open models and infrastructure, evaluation, multilingual NLP, and AI for science. Biderman was the second author of The Pile, an 800GB dataset of diverse English text for language modeling posted at the end of 2020, and of the 2022 paper describing GPT-NeoX-20B, an open-source autoregressive language model.¹⁵⁶

In 2023 Biderman was first author of the paper on the Pythia suite, built to let researchers study how language models change across training and scale. More recently Biderman co-authored the Common Pile v0.1, an 8TB dataset of public domain and openly licensed text, in 2025.²³⁷

Evaluation

Biderman is among the credited authors of the Language Model Evaluation Harness and first author of Lessons from the Trenches on Reproducible Evaluation of Language Models (2024), which draws on experience maintaining that framework to discuss reproducible evaluation practice.⁴⁸

Work and contributions

• Pythia

  First author of the Pythia paper and first-listed author in the repository's citation for the model suite.²³

• LM Evaluation Harness

  Listed among the authors in the project's citation entry for the evaluation framework.⁴

• The Pile Open data

  Second author, after Leo Gao, of the 2020 paper introducing The Pile, an 800GB dataset of diverse text for language modeling.⁵

• GPT-NeoX-20B Open weights

  Second author, after Sid Black, of the 2022 paper describing GPT-NeoX-20B, an open-source autoregressive language model.⁶

• Pythia 12B Open weights

  First author of the Pythia paper, which the Pythia 12B model card cites.²

Timeline

1. Dec 31, 2020Co-authored The Pile, an 800GB dataset of diverse text for language modeling.⁵
2. Apr 14, 2022Co-authored GPT-NeoX-20B: An Open-Source Autoregressive Language Model.⁶
3. Apr 3, 2023Posted the Pythia paper as first author, introducing a suite of models for studying training and scaling.²
4. May 23, 2024Posted Lessons from the Trenches on Reproducible Evaluation of Language Models as first author.⁸
5. Jun 5, 2025Co-authored The Common Pile v0.1, a dataset of public domain and openly licensed text.⁷

Where they work

• EleutherAI 5 people in the wiki

Related people

• Leo GaoAlso at EleutherAI
• Sid BlackAlso at EleutherAI
• Connor LeahyAlso at EleutherAI
• Katherine CrowsonAlso at EleutherAI
• Diogo AlmeidaLanguage models · Evaluation and benchmarks
• Aylin CaliskanEvaluation and benchmarks · Language models
• Tatsunori HashimotoLanguage models · Evaluation and benchmarks
• Pradeep DasigiLanguage models · Evaluation and benchmarks
• Ruchir PuriLanguage models · Evaluation and benchmarks
• Eric Michael SmithEvaluation and benchmarks · Language models