Uses and Limitations

Intended Use

The primary intended use of Pythia is research on the behavior, functionality, and limitations of large language models. This suite is intended to provide a controlled setting for performing scientific experiments. We also provide 154 checkpoints per model: initial step0, 10 log-spaced checkpoints step{1,2,4...512}, and 143 evenly-spaced checkpoints from step1000 to step143000. These checkpoints are hosted on Hugging Face as branches. Note that branch 143000 corresponds exactly to the model checkpoint on the main branch of each model.

You may also further fine-tune and adapt Pythia-160M-deduped for deployment, as long as your use is in accordance with the Apache 2.0 license. Pythia models work with the Hugging Face Transformers Library. If you decide to use pre-trained Pythia-160M-deduped as a basis for your fine-tuned model, please conduct your own risk and bias assessment.

Out-of-scope use

The Pythia Suite is not intended for deployment. It is not a in itself a product and cannot be used for human-facing interactions. For example, the model may generate harmful or offensive text. Please evaluate the risks associated with your particular use case.

Pythia models are English-language only, and are not suitable for translation or generating text in other languages.

Pythia-160M-deduped has not been fine-tuned for downstream contexts in which language models are commonly deployed, such as writing genre prose, or commercial chatbots. This means Pythia-160M-deduped will not respond to a given prompt the way a product like ChatGPT does. This is because, unlike this model, ChatGPT was fine-tuned using methods such as Reinforcement Learning from Human Feedback (RLHF) to better “follow” human instructions.

Limitations and biases

The core functionality of a large language model is to take a string of text and predict the next token. The token used by the model need not produce the most “accurate” text. Never rely on Pythia-160M-deduped to produce factually accurate output.

This model was trained on the Pile, a dataset known to contain profanity and texts that are lewd or otherwise offensive. See Section 6 of the Pile paper for a discussion of documented biases with regards to gender, religion, and race. Pythia-160M-deduped may produce socially unacceptable or undesirable text, even if the prompt itself does not include anything explicitly offensive.

If you plan on using text generated through, for example, the Hosted Inference API, we recommend having a human curate the outputs of this language model before presenting it to other people. Please inform your audience that the text was generated by Pythia-160M-deduped.

Quickstart

Pythia models can be loaded and used via the following code, demonstrated here for the third pythia-70m-deduped checkpoint:

from transformers import GPTNeoXForCausalLM, AutoTokenizer

model = GPTNeoXForCausalLM.from_pretrained(
  "EleutherAI/pythia-70m-deduped",
  revision="step3000",
  cache_dir="./pythia-70m-deduped/step3000",
)

tokenizer = AutoTokenizer.from_pretrained(
  "EleutherAI/pythia-70m-deduped",
  revision="step3000",
  cache_dir="./pythia-70m-deduped/step3000",
)

inputs = tokenizer("Hello, I am", return_tensors="pt")
tokens = model.generate(**inputs)
tokenizer.decode(tokens[0])

Revision/branch step143000 corresponds exactly to the model checkpoint on the main branch of each model. For more information on how to use all Pythia models, see documentation on GitHub.

Training

Training data

Pythia-160M-deduped was trained on the Pile after the dataset has been globally deduplicated. The Pile is a 825GiB general-purpose dataset in English. It was created by EleutherAI specifically for training large language models. It contains texts from 22 diverse sources, roughly broken down into five categories: academic writing (e.g. arXiv), internet (e.g. CommonCrawl), prose (e.g. Project Gutenberg), dialogue (e.g. YouTube subtitles), and miscellaneous (e.g. GitHub, Enron Emails). See the Pile paper for a breakdown of all data sources, methodology, and a discussion of ethical implications. Consult the datasheet for more detailed documentation about the Pile and its component datasets. The Pile can be downloaded from the official website, or from a community mirror.

Training procedure

All models were trained on the exact same data, in the exact same order. Each model saw 299,892,736,000 tokens during training, and 143 checkpoints for each model are saved every 2,097,152,000 tokens, spaced evenly throughout training, from step1000 to step143000 (which is the same as main). In addition, we also provide frequent early checkpoints: step0 and step{1,2,4...512}. This corresponds to training for just under 1 epoch on the Pile for non-deduplicated models, and about 1.5 epochs on the deduplicated Pile.

All Pythia models trained for 143000 steps at a batch size of 2M (2,097,152 tokens). See GitHub for more details on training procedure, including how to reproduce it. Pythia uses the same tokenizer as GPT-NeoX- 20B.

Evaluations

All 16 Pythia models were evaluated using the LM Evaluation Harness. You can access the results by model and step at results/json/* in the GitHub repository. Expand the sections below to see plots of evaluation results for all Pythia and Pythia-deduped models compared with OPT and BLOOM.

Changelog

This section compares differences between previously released Pythia v0 and the current models. See Appendix B of the Pythia paper for further discussion of these changes and the motivation behind them. We found that retraining Pythia had no impact on benchmark performance.

• All model sizes are now trained with uniform batch size of 2M tokens. Previously, the models of size 160M, 410M, and 1.4B parameters were trained with batch sizes of 4M tokens.
• We added checkpoints at initialization (step 0) and steps {1,2,4,8,16,32,64, 128,256,512} in addition to every 1000 training steps.
• Flash Attention was used in the new retrained suite.
• We remedied a minor inconsistency that existed in the original suite: all models of size 2.8B parameters or smaller had a learning rate (LR) schedule which decayed to a minimum LR of 10% the starting LR rate, but the 6.9B and 12B models all used an LR schedule which decayed to a minimum LR of 0. In the redone training runs, we rectified this inconsistency: all models now were trained with LR decaying to a minimum of 0.1× their maximum LR.

Naming convention and parameter count

Pythia models were renamed in January 2023. It is possible that the old naming convention still persists in some documentation by accident. The current naming convention (70M, 160M, etc.) is based on total parameter count.