Pythia 1B
Pythia 1B is a base language model in a suite designed for controlled training research. This record describes the current standard-Pile weights.
Research release
Architecture and scale
The 16-layer architecture has a hidden width of 2,048 and eight attention heads. The current standard-data run used bfloat16 because the authors could not resolve late loss spikes with float16. That training difference matters for numerical replication.pythia-1b model cardpythia-1b model configurationPythia releases, corrections and checkpoint structure
Weights and training revisions
The release includes 154 saved training steps. The data order is shared with other sizes in the same suite. The main branch provides the final model; a step revision selects an earlier checkpoint. This base model has not been trained as a conversational assistant.pythia-1b model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Launch date | |
| Access | Downloadable |
| Model identifiers | EleutherAI/pythia-1b |
| Context window | 2,048 tokens Maximum position count in the released model configuration. pythia-1b model configuration |
| Nominal parameters | 1B including embeddingspythia-1b model card |
| Network | 16 layers; hidden width 2,048; 8 attention headspythia-1b model configuration |
| Training data | Standard Pilepythia-1b model card |
| License | Apache 2.0pythia-1b model card |
Related models
- Pythia 1B Deduped
EleutherAI ยท Pythia
A 1B Pythia checkpoint trained on deduplicated Pile text.
Launch date Research releaseDownloadable