Research map / Large language models / Language, safety and interpretability
Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling
Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, Aviya Skowron, Lintang Sutawika and 1 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on the near-deduplicated Pile instead of the original Pile (ablation) | Pythia · Language modeling benchmarks | evaluation performance | No clear effect |
| Use parallel attention and feedforward sublayers at all model scales | Pythia · NLP benchmarks | performance relative to OPT | No clear effect |
| Use batch sizes 4× to 8× standard for models under 1 billion parameters | Pythia · Training runs for models under 1 billion parameters | convergence | No clear effect |
| Use larger batch sizes for smaller models | Pythia · Training time for smaller models | wall-clock speed-up: 10$\times$ | Improved |
| Replace masculine pronouns with feminine counterparts in the final 7% or 21% of training | Pythia-deduped · WinoBias | stereotypical accuracy | Improved |
| Replace masculine pronouns with feminine counterparts in the final 7% or 21% of training | Pythia-deduped · CrowS-Pairs gender | % Stereotype | Improved |
| Replace masculine pronouns with feminine counterparts in the final 7% or 21% of training | Pythia-deduped · LAMBADA | perplexity | Improved |
| Place sequences at different positions in training | Pythia · Training corpus memorization | memorized-sequence occurrence rate | No clear effect |
| Increase model size to 2.8 billion parameters or more and continue training past 65,000 steps | Pythia-deduped · Arithmetic and TriviaQA | correlation of accuracy with relevant term frequency: 65,000 training steps | Improved |
| Train on a second epoch of the deduplicated Pile | Pythia-deduped · Language-modeling benchmarks | evaluation scores | No clear effect |
Later papers that compared with this method
2 later accepted papers report Pythia-12B as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- PonderLM: Pretraining Language Models to Ponder in Continuous Space (ICLR 2026 · better on 0 of 2 shared results)
- OpenAssistant Conversations - Democratizing Large Language Model Alignment (NeurIPS 2023 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- PolyPythias: Stability and Outliers across Fifty Language Model Pre-Training Runs (ICLR 2025 · similarity 0.92)
- Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.91)
- Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning (ICLR 2024 · similarity 0.90)
- LLMs on the Line: Data Determines Loss-to-Loss Scaling Laws (ICML 2025 · similarity 0.90)
- A Probabilistic Perspective on Unlearning and Alignment for Large Language Models (ICLR 2025 · similarity 0.90)
- Scaling Laws with Vocabulary: Larger Models Deserve Larger Vocabularies (NeurIPS 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
