Pythia (model suite)
Sparse Autoencoders Find Highly Interpretable Features in Language Models — inherited
A suite of open, reproducible language models of varying sizes (Biderman et al., 2023). This paper trains sparse autoencoders on the residual streams of Pythia-70M (din=512) and Pythia-410M (din=1024), its primary experimental subjects throughout.
None of the interpretability claims in this paper are really about the language model they're demonstrated on — every learned feature, every automated interpretability score, every patching result belongs to a small, openly released model chosen specifically because its weights, training data, and training order are all public, not because it's the most capable model available.
The page covers what makes this model suite distinctive as a research tool, full parameter access and dense checkpointing across a range of sizes, what gets trained on it specifically in this paper, and why an open, reproducible substrate matters for a method whose entire premise is inspecting a model's internals rather than just its outputs.
The default experimental substrate, not a subject of study
Pythia is not something this paper investigates; it is what every other technique in the paper is run on. Every dictionary feature, every autointerpretability score, every patching result, and every case-study feature described elsewhere on this wiki is a fact about a sparse autoencoder trained on a Pythia model's residual stream -- it belongs to open-models-as-interpretability-testbeds because its entire purpose in the corpus is to be a stable, fully-open target for interpretability tooling rather than a capability to be improved.
Built for exactly this kind of research
Pythia (Biderman, Schoelkopf, Anthony, Bradley, O'Brien, Hallahan, Khan, Purohit, Prashanth, Raff et al., 2023) is a suite of decoder-only transformer language models released by EleutherAI spanning roughly 70 million to 12 billion parameters, trained on identical data in an identical order across the whole suite, with 154 intermediate checkpoints released per model. That design -- full parameter access, reproducible training order, and dense checkpointing -- was deliberately built to support scaling and interpretability research that needs controlled comparisons across model size and training time, precisely the property that makes it a natural substrate for a method whose entire claim rests on inspecting a model's internals directly.
What this paper trains on it
This paper trains its sparse autoencoders mainly on the residual streams of Pythia-70M ($d_{in}=512$) and Pythia-410M ($d_{in}=1024$) (sparse-autoencoders, §"2 Taking features out of superposition with sparse dictionary learning", p. 2). Pythia-410M's layer 11 supplies the headline IOI-patching results in Figure 3, with Appendix F extending the same procedure across layers 3, 7, 11, 15, 19, and 23 of its 24 total layers; Pythia-70M's layer-2 residual stream is separately used for the reconstruction-fidelity check that finds substituting a layer's activations with their dictionary reconstruction raises perplexity on the Pile from 25 to 40 (sparse-autoencoders, §"6.2 Limitations and future work", p. 9).
Why a small open model, not a frontier one
Choosing Pythia over a larger closed model is itself part of the paper's method, not incidental: because Pythia's weights and training data are fully public, the paper can freely publish activation datasets, feature dictionaries, and case-study details -- such as unembedding feature 52027 to read off its top tokens -- without any dependence on privileged API access. That openness dovetails with the paper's own stated ambition for its method to serve as "a foundation for future mechanistic interpretability work" (sparse-autoencoders, §"Abstract", p. 1): reproducibility down to the substrate is a precondition for that kind of foundation.