This repository is for EleutherAI's project Pythia which combines interpretability analysis and scaling laws to understand how knowledge develops and evolves during training in autoregressive transformers. For detailed info on the models, their training, and their properties, please see our paper Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling.

The Pythia suite was developed with the explicit purpose of enabling research in interpretability, learning dynamics, and ethics and transparency for which existing model suites were inadequate. The key features of the Pythia suite are:

1. All models, data, and code used in the paper are publicly released, enabling full reproducibility of results. All results in our paper have been independently verified by at least one other lab.
2. All models feature 154 checkpoints saved throughout training, enabling the study of learning dynamics of LLMs.
3. All models were trained on the same data in the same order, enabling researchers to explore causal interventions on the training process.

Aside from the Pythia suite itself, this repository also acts as a hub containing information, code, and reproducibility instructions for the following papers:

• Emergent and Predictable Memorization in Large Language Models code paper

• Pythia: Interpreting Transformers Across Time and Scale
  • Models
  • Changelog
• Using Pythia
  • Quickstart
  • Reproducing Training
  • Exploring the Dataset
  • Pythia Paper Replication
• Benchmark Scores
• Research Building on Pythia
  • Interpretability Research
  • Learning Dynamics Research
  • Ethics and Transparency Research
  • Other Notable Research
• Citation Details
• License

We train and release a suite of 8 model sizes on the the Pile (paper, datasheet) as well as the Pile with deduplication applied. All 8 model sizes are trained on the exact same data, in the exact same order. Each model saw 299,892,736,000 ~= 299.9B tokens during training. This corresponds to just under 1 epoch on the Pile for non-"deduped" models, and ~= 1.5 epochs on the deduped Pile (which contains 207B tokens in 1 epoch). All models are trained with mixed precision, using fp16 for all models except EleutherAI/pythia-1b which trained with bf16, because in fp16 the model experienced an irreconcilable loss spike late in training.

To promote research on the learning dynamics of LLMs we make 154 checkpoints available for each model, representing steps 0 (initialization), 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1000, and then every 1,000 subsequent steps.

Config files used to train these models with the GPT-NeoX library can be found at the models/ directory within this repository, as well as in the GPT-NeoX library itself.

We also upload the pre-tokenized data files and a script to reconstruct the dataloader as seen during training for all models. See Reproducing Training section for more details.

[November 2, 2023] We have added 14M and 31M models at the request of some researchers. We plan on training deduped versions of these models in the future.

[April 3, 2023] We have released a new version of all Pythia models, fixing various inconsistencies in the original suite. Please see our paper for details on the changes. The old models ("v0") remain available here and may be useful for ablation studies.

[January 20, 2023] On January 20, 2023, we chose to rename the Pythia model suite to include both embedding layer and unembedding layer parameters in our total parameter counts, in line with many other model suites and because we believe this convention better reflects the on-device memory usage of these models. We also discovered that due to a typo one of our models was smaller than we thought, and replaced it with a model of the intended size. See here for more details.

All Pythia models are hosted on the Huggingface hub. They can be loaded and used via the following code (shown for the 3000-step pythia-70M-deduped model 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])

All models were trained for the equivalent of 143000 steps at a batch size of 2,097,152 tokens. Revision/branch step143000 corresponds exactly to the model checkpoint on the main branch of each model.

We additionally have all model checkpoints in the format accepted by the GPT-NeoX library, with final-step checkpoints+optimizer states downloadable from the Hugging Face Hub at EleutherAI/neox-ckpt-pythia-xxx-deduped-v1 but do not serve them for all steps at scale due to size of optimizer states and anticipated lower demand. If you would like to perform analysis using the intermediate models within the GPT-NeoX codebase, or would like the optimizer states for other steps, please email [email protected] and [email protected].

❗ pythia-{size}-v0 models on Huggingface of sizes 160m, 410m, 1.4b were trained with a batch size of 4M tokens across 71500 steps and checkpointed every 500 steps. The step names on Huggingface for these v0 models are renamed for consistency with all 2M batch models so the model checkpointed labeled step1000 of pythia-1.4b-v0 was actually step 500, but has seen the same number of tokens as the other step1000 checkpoints.

(Expanded reproduction instructions provided by @BaruchG ).

We provide the training data for replication of our training runs. The GPT-NeoX library requires the pre-tokenized training data in the form of 2 memory-mapped numpy arrays: a .bin and .idx file. We provide these files via the Hugging Face hub. To download and use the deduplicated Pile training data:

git lfs clone https://huggingface.co/datasets/EleutherAI/pythia_deduped_pile_idxmaps

# Optionally, to ensure against corrupt files
python utils/checksum_shards.py

python utils/unshard_memmap.py --input_file ./pythia_deduped_pile_idxmaps/pile_0.87_deduped_text_document-00000-of-00082.bin --num_shards 83 --output_dir ./pythia_pile_idxmaps/

# The correct sha256 for the full file is 0cd548efd15974d5cca78f9baddbd59220ca675535dcfc0c350087c79f504693
# This can be checked with sha256sum ./pythia_pile_idxmaps/*

This will take over a day to run, though it should not require more than 5 GB of RAM. We recommend downloading this rather than retokenizing the Pile from scratch in order to guarantee preservation of the data order seen by the Pythia models. In addition to the training data, you will need to make a local copy of the tokenizer we used to train our models. You can find it here.

Next you will need to set up the training environment:

git clone https://github.com/EleutherAI/gpt-neox.git
cd gpt-neox
git checkout v1.0
pip install -r requirements/requirements-flashattention.txt
wget https://github.com/EleutherAI/pythia/blob/main/models/160M/pythia-160m-deduped.yml
docker build -t pythia:latest .

After the container finishes building, run the container using the following command (from the root of the GPT-NeoX repo with your pythia yaml accessible from within that folder):

docker run --runtime=nvidia --rm -it -e NVIDIA_VISIBLE_DEVICES=0,1,2,3 --shm-size=1g --ulimit memlock=-1 --mount type=bind,src=$PWD,dst=/gpt-neox -v $(pwd):/workspace/ pythia:latest bash

You can use the -v argument to add more connected volumes for the dataset and the Yaml file if is not accessible from within the docker container.

Change the lines of the data paths and tokenizer paths as follows:

  "train-data-paths": ["/fsx/pile/pile_20B_tokenizer_text_document"], #point this to your folder which was generated in step 1 containing the .bin and .idx file
  "valid-data-paths": ["/fsx/pile/pile_20B_tokenizer_text_document"], #point this to your folder which was generated in step 1 containing the .bin and .idx file
  "test-data-paths": ["/fsx/pile/pile_20B_tokenizer_text_document"], #point this to your folder which was generated in step 1 containing the .bin and .idx file

  "tokenizer-type": "HFTokenizer",
  "vocab-file": "/fsx/pile/20B_tokenizer.json", # point this to the tokenizer retrieved in step 2

Depending on how much VRAM you have available you may need to adjust the batch sizes. The total batch size is calculated via Total GPUs * train_micro_batch_size_per_gpu * gradient_accumulation_steps / (pipe-parallel-size * model-parallel-size) and needs to be kept at 1024 to match the Pythia training batch size. You

   "train_micro_batch_size_per_gpu": XXX, # make this a value that will fit within your GPU memory
   "gradient_accumulation_steps": 1, # make this a value to compensate to make the total batch size 1024.

If you would like your weights to be saved add that information to the yaml file as well. For example, to save in the checkpoints folder, at the bottom you can add:

  "launcher": "slurm",
  "deepspeed_slurm": false,

  "save": "checkpoints",
  "load": "checkpoints",
  "checkpoint_validation_with_forward_pass": False,
}

Make sure the paths are the paths from inside your docker container and if you want the weights to have persistence, make sure that they are accessible from outside the container, for example in /workspace/ .

You should now be able to start training your model by running:

python deepy.py train.py pythia-160m-deduped.yml  2>&1 | tee output.txt

the output will be saved to output.txt, if you don’t want that just delete the end.

In order to convert your model to the Hugging Face transformers format, you can use the script tools/convert_to_hf.py from within the GPT-NeoX library. You may have to add from typing import List to the type of the file and change the line here from list[torch.Tensor] to List[torch.Tensor]. You can then run the script like this to convert the weights at step 143000:

python tools/convert_to_hf.py --input_dir checkpoints/global_step143000/ --config_file checkpoints2/global_step 143000/configs/pythia-70m.yml --output_dir ./output/ 

This should output a file structure similar to the one found at https://huggingface.co/EleutherAI/pythia-70m-deduped/tree/main.

❗ Sometimes people find that they don't end up with the right tokenizer for reasons we have been unable to debug. If your tokenizer_config.json looks different than the one here and special_tokens_map.json look different than here you may need to replace them with the ones on Huggingface.

To run evaluations using our evaluation library, install the containers here (tested with the 4.28 and 4.29 versions). After setting up that docker container run:

git clone https://github.com/EleutherAI/lm-evaluation-harness
cd lm-evaluation-harness
pip install -e .

as outlined in the Harness repository. You should then be able to run the benchmark by pointing it at your weights (which should be in your container) by running a command similar to this:

python3 main.py --model hf-causal-experimental  --model_args pretrained=../gpt-neox/output/ --tasks lambada_openai,piqa,winogrande,arc_easy,sciq,wikitext --device cuda:0

We provide a tool to view particular portions of the training dataloader used by all models during training, at utils/batch_viewer.py.

First, we need to clone the Pythia repository:

git clone https://github.com/EleutherAI/pythia

Next, we must install dependencies:

pip install torch==1.13.0+cu117 -f https://download.pytorch.org/whl/torch/
pip install numpy tqdm huggingface_hub

Next, we must download the appropriate dataset. We provide preshuffled versions of the duped and deduped pile. Download the appropriate one using Huggingface's utilities as follows:

Tip: Make sure to replace path/to/* to appropriate paths where you intend to save datasets downloaded from Huggingface.

• To download standard version, use

  from huggingface_hub import hf_hub_download
  hf_hub_download(repo_id="EleutherAI/pile-standard-pythia-preshuffled", repo_type="dataset", cache_dir="path/to/local/folder")

• To download the deduped version, use

  from huggingface_hub import hf_hub_download
  hf_hub_download(repo_id="EleutherAI/pile-standard-pythia-preshuffled", repo_type="dataset", cache_dir="path/to/local/folder")

You can now merge the files by using the script utils/unshard_mmap.py :

python3 utils/unshard_mmap.py --input_file "path/to/local/folder/document-00000-of-00020.bin" --num_shards 21 --output_dir "path/to/merged/folder/"

Make sure to also copy index file to the merged folder, using the command

cp path/to/local/folder/document.idx path/to/merged/folder/document.idx

Now, we're all set up to run utils/batch_viewer.py !

python3 utils/batch_viewer.py \
  --start_iteration 0 \
  --end_iteration 1000 \
  --load_path path/to/merged/folder/document \
  --save_path path/to/save/folder/ \
  --conf_dir utils/dummy_config.yml 

This will save a separate file containing all the indicies as a numpy array.

You can now load this using numpy as

import numpy as np

indicies = np.load("path/to/save/folder/indicies.npy")

These indicies contain tokenized sequences of integers of size (None, 2049), where an integer corresponds to a unique token index. Note that documents are concatenated and saperated by an EOD token. Thus, each sample or batch may not start with an EOD token. During training, target tokens are left shifted by 1. Thus, a model of sequence length 2048 requires 2049 length sequences for training (For more info, refer to this comment)

We provide further information for those interested in replicating our case studies performed in the Pythia suite paper in the case-studies/ folder of this repository, being

• Memorization density over training
• Intervention on pronoun frequencies in pretraining
• Term frequency effects over training

We also provide benchmark 0-shot and 5-shot results on a variety of NLP datasets:

• Lambada (lambada_openai)
• Wikitext (wikitext)
• PiQA (piqa)
• SciQ (sciq)
• WSC (wsc)
• Winogrande (winogrande)
• ARC-challenge (arc_challenge)
• ARC-easy (arc_easy)
• LogiQA (logiqa)
• BLiMP (blimp_*)
• MMLU (hendrycksTest*)

Evaluations were performed in GPT-NeoX using the LM Evaluation Harness, and are viewable by model and step at evals/pythia-v1/*/* in this repository.

Our primary goal with the Pythia project is to enable research on interpretability and learning dynamics at EleutherAI and in the community writ large. Here we document select papers using our models, focusing on work that is uniquely empowered by the Pythia suite. For a complete list of papers citing Pythia, see here.

• Belrose, et al. "Eliciting latent predictions from transformers with the tuned lens." arXiv preprint arXiv:2303.08112 (2023).
• Brown, et al. "Understanding the Inner Workings of Language Models Through Representation Dissimilarity." arXiv preprint arXiv:2310.1499 (2023).
• Feng and Steinhardt. "How do Language Models Bind Entities in Context?." arXiv preprint arXiv:2310.17191 (2023).
• Garde, Kran, and Barez. "DeepDecipher: Accessing and Investigating Neuron Activation in Large Language Models." arXiv preprint arXiv:2310.01870 (2023).
• Gurnee, et al. "Finding Neurons in a Haystack: Case Studies with Sparse Probing." arXiv preprint arXiv:2305.01610 (2023).
• Stolfo, Belinkov, and Sachan. "Understanding Arithmetic Reasoning in Language Models using Causal Mediation Analysis." arXiv preprint arXiv:2305.15054 (2023).

• Gupta, et al. "Continual Pre-Training of Large Language Models: How to re-warm your model?." Workshop on Efficient Systems for Foundation Models @ ICML (2023).
• Michaelov and Bergen. "Emergent inabilities? Inverse scaling over the course of pretraining." arXiv preprint arXiv:2305.14681 (2023).
• Roger. "Large Language Models Sometimes Generate Purely Negatively-Reinforced Text." arXiv preprint arXiv:2306.07567 (2023).
• Sanyal, et al. "Understanding the Effectiveness of Early Weight Averaging for Training Large Language Models." arXiv preprint arXiv:2306.03241 (2023).
• Tian, et al. "JoMA: Demystifying Multilayer Transformers via JOint Dynamics of MLP and Attention." arXiv preprint arXiv:2310.0053 (2023).

• Biderman, et al. "Emergent and predictable memorization in large language models." NeurIPS (2023).
• Choi, Shavit, and Duvenaud. "Tools for Verifying Neural Models' Training Data." arXiv preprint arXiv:2307.00682 (2023).
• Ippolito, et al. "Reverse-Engineering Decoding Strategies Given Blackbox Access to a Language Generation System." Proceedings of the 16th International Natural Language Generation Conference. 2023.
• Li et al. "MoPe: Model Perturbation-based Privacy Attacks on Language Models." Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (2023).
• Min et al. "SILO Language Models: Isolating Legal Risk In a Nonparametric Datastore." arXiv preprint arXiv:2308.0443 (2023).

• Sileo and Lernould. "Mindgames: Targeting theory of mind in large language models with dynamic epistemic modal logic." arXiv preprint arXiv:2305.03353 (2023).
• Köpf, et al. "OpenAssistant Conversations--Democratizing Large Language Model Alignment." arXiv preprint arXiv:2304.07327 (2023).
• Ye, et al. "Language Versatilists vs. Specialists: An Empirical Revisiting on Multilingual Transfer Ability." arXiv preprint arXiv:2306.06688 (2023).

If you use the Pythia models in your research, please cite our paper via:

@inproceedings{biderman2023pythia,
  title={Pythia: A suite for analyzing large language models across training and scaling},
  author={Biderman, Stella and Schoelkopf, Hailey and Anthony, Quentin Gregory and Bradley, Herbie and O’Brien, Kyle and Hallahan, Eric and Khan, Mohammad Aflah and Purohit, Shivanshu and Prashanth, USVSN Sai and Raff, Edward and others},
  booktitle={International Conference on Machine Learning},
  pages={2397--2430},
  year={2023},
  organization={PMLR}
}

If you use data or results from other papers found in this repository, please cite the corresponding papers. Citation information can be found in the respective README and are also reproduced below for convenience:

@inproceedings{biderman2023emergent,
      title={Emergent and Predictable Memorization in Large Language Models}, 
      author={Biderman, Stella and Prashanth, USVSN Sai and Sutawika, Lintang and Schoelkopf, Hailey and Anthony, Quentin and Purohit, Shivanshu and Raff, Edward},
      journal={Advances in Neural Information Processing Systems},
      year={2023}
}

If you are interested in citing our training data, training library, or evaluation library you can do so with the following:

@article{gao2020pile,
  title={The pile: An 800gb dataset of diverse text for language modeling},
  author={Gao, Leo and Biderman, Stella and Black, Sid and Golding, Laurence and Hoppe, Travis and Foster, Charles and Phang, Jason and He, Horace and Thite, Anish and Nabeshima, Noa and others},
  journal={arXiv preprint arXiv:2101.00027},
  year={2020}
}

@article{biderman2022datasheet,
  title={Datasheet for the pile},
  author={Biderman, Stella and Bicheno, Kieran and Gao, Leo},
  journal={arXiv preprint arXiv:2201.07311},
  year={2022}
}

@software{gpt-neox-library,
  title = {{GPT-NeoX: Large Scale Autoregressive Language Modeling in PyTorch}},
  author = {Andonian, Alex and Anthony, Quentin and Biderman, Stella and Black, Sid and Gali, Preetham and Gao, Leo and Hallahan, Eric and Levy-Kramer, Josh and Leahy, Connor and Nestler, Lucas and Parker, Kip and Pieler, Michael and Phang, Jason and Purohit, Shivanshu and Schoelkopf, Hailey and Stander, Dashiell and Songz, Tri and Tigges, Curt and Thérien, Benjamin and Wang, Phil and Weinbach, Samuel},
  url = {https://www.github.com/eleutherai/gpt-neox},
  doi = {10.5281/zenodo.5879544},
  month = {9},
  year = {2023},
  version = {2.0.0},
}

@misc{eval-harness,
  author       = {Gao, Leo and Tow, Jonathan and Abbasi, Baber and Biderman, Stella and Black, Sid and DiPofi, Anthony and Foster, Charles and Golding, Laurence and Hsu, Jeffrey and Le Noac'h, Alain and Li, Haonan and McDonell, Kyle and Muennighoff, Niklas and Ociepa, Chris and Phang, Jason and Reynolds, Laria and Schoelkopf, Hailey and Skowron, Aviya and Sutawika, Lintang and Tang, Eric and Thite, Anish and Wang, Ben and Wang, Kevin and Zou, Andy},
  title        = {A framework for few-shot language model evaluation},
  month        = sep,
  year         = 2021,
  publisher    = {Zenodo},
  version      = {v0.0.1},
  doi          = {10.5281/zenodo.5371628},
  url          = {https://doi.org/10.5281/zenodo.5371628}
}

The following license applies to all code in this GitHub repo, as well as the Pythia models and any other copyrightable artifacts contained in this repository.

   Copyright 2023 EleutherAI

   Licensed under the Apache License, Version 2.0 (the "License");
   you may not use this file except in compliance with the License.
   You may obtain a copy of the License at

       https://www.apache.org/licenses/LICENSE-2.0

   Unless required by applicable law or agreed to in writing, software
   distributed under the License is distributed on an "AS IS" BASIS,
   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
   See the License for the specific language governing permissions and
   limitations under the License.