AI Potluck
Model components / Base / pretrained models

Pythia

EleutherAI

EleutherAI's suite of sixteen models from 70M to 12B parameters, eight sizes trained twice on The Pile and its deduplicated variant, in the same data order, with 154 intermediate checkpoints released for each. It is a scientific instrument for studying how language models learn rather than a production family, and the dataloader tooling makes the exact token order reconstructable.

These are research instruments rather than models anyone deploys, which is what the capability axis reflects. Verified 2026-08-13 via the Pythia paper, the Pile dataset entry, the EleutherAI/pythia repository and the pythia-12b model card.

Openness

5 high confidence
5.0
weights
open(Apache-2.0)
data
open(the Pile, with dataloader reconstruction tools)
code
open(training+analysis code, GitHub)
checkpoints
open(154 per model)
license
Apache-2.0(OSI)

The reference case for a full open-source score in this category. Apache-2.0 weights, the Pile released as the training corpus together with dataloader tools that reconstruct the exact token order, training and analysis code on GitHub, and 154 intermediate checkpoints per model size. Every dimension this axis asks about is answered by something downloadable, which is what separates a 5 from the open-weights tier: the training run is reproducible, not merely the model usable.

  • https://huggingface.co/api/datasets/EleutherAI/pile recorded 2026-08-13

    `gated: false`, `private: false`, 2,308 downloads, 502 likes. The Pile is published and ungated on the Hub.

  • https://arxiv.org/abs/2304.01373 recorded 2026-08-13

    Abstract: "We provide public access to 154 checkpoints for each one of the 16 models, alongside tools to download and reconstruct their exact training dataloaders" - which is both the checkpoint claim and the reconstruction tooling that makes the corpus reproducible in token order.

  • https://github.com/EleutherAI/pythia recorded 2026-08-13

    Repository page, described as "The hub for EleutherAI's work on interpretability and learning dynamics" - training and analysis code, public.

  • https://huggingface.co/EleutherAI/pythia-12b/blob/main/README.md recorded 2026-08-13

    Model card front matter: `license: apache-2.0`, `datasets: [EleutherAI/pile]`; License apache-2.0 in the repo sidebar.

Adoption

2 high confidence
2.0

30,007 downloads in the trailing 30 days for the single declared artifact EleutherAI/pythia-12b, which bands at level 2 (10K-100K) on the model adoption scale.

Capability

1 high confidence
1.0

Not built for capability. Pythia models (max 12B, 2023-era Pile training) are intentionally research instruments for studying scaling/learning dynamics, not competitive on modern capability benchmarks. Scored 1 on a capability axis where the comparison set is 2026 frontier models.

  • https://arxiv.org/abs/2304.01373 recorded 2026-08-13

    "we introduce Pythia, a suite of 16 LLMs all trained on public data seen in the exact same order and ranging in size from 70M to 12B parameters"; the presented results are case studies in memorization, term frequency and bias, not capability benchmarks.

Verified 2026-08-13