This repository is for EleutherAI's work-in-progress project Pythia which combines interpretability analysis and scaling laws to understand how knowledge develops and evolves during training in autoregressive transformers.

We train and release a suite of 8 model sizes on 2 different datasets: the Pile, 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, and 143 checkpoints for each model are saved every 2,097,152,000 ~= 2B tokens, evenly spaced throughout 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).

Config files used to train these models within the GPT-NeoX library can be found at the models/ directory within this repository.

We are planning on releasing a user-friendly utility, but the dataset loader in GPT-NeoX (which is the same as the Megatron-DS one, AFAIK) allows you to save and export the seed used to shuffle data. If you just build the GPT2Dataset with the right neox args (provided with the model checkpoints) then it should be easily accessible — just grab the [(BS * starting iter) + index]th element of the dataset object.

All Pythia models are hosted on the Huggingface hub. They can be loaded and used via the following code (shown for the 3rd pythia-19M-deduped model checkpoint):

from transformers import GPTNeoXForCausalLM

model = GPTNeoXForCausalLM.from_pretrained(
  "EleutherAI/pythia-19m-deduped",
  revision="step3000",
  cache_dir="./pythia-19m-deduped/step3000",
)

tokenizer = AutoTokenizer.from_pretrained(
  "EleutherAI/pythia-19m-deduped",
  revision="step3000",
  cache_dir="./pythia-19m-deduped/step3000",
)

inputs = tokenizer("Hello, I am", return_tensors="pt")
model.generate(**inputs)

All models were trained for the equivalent of 143000 steps at a batch size of 2,097,152 tokens. Revision/branch step143000 (e.g. https://huggingface.co/EleutherAI/pythia-19m-deduped/tree/step143000) corresponds exactly to the model checkpoint on the main branch of each model.

Models with a batch size of 4M tokens listed were originally trained for 71500 steps instead, and checkpointed every 500 steps. The checkpoints on Huggingface are renamed for consistency with all 2M batch models, so step1000 is the first checkpoint for the 1.3B that was saved (corresponding to step 500 in training), and step1000 is likewise the first 6.7B checkpoint that was saved (corresponding to 1000 "actual" steps.)

We additionally have all model checkpoints in the format accepted by the GPT-NeoX library, but do not serve them at scale due to size of optimizer states and anticipated lower demand. If you would like to perform analysis using the models within the GPT-NeoX codebase, or would like the optimizer states, please email us at [email protected] to arrange access.

A common explanation for language model training dynamics is that LMs have a mass of knowledge and when they come across new information they glom that knowledge on and slowly integrate it into the mass over time. One prediction that this mental model makes is that tokens encountered later in training will be more likely to be memorized than ones encountered earlier in training, as the model will not have time to adjust its representations to store the info without memorization. The primary goal of this experiment is to disprove this prediction and demonstrate that training order doesn't influence memorization.