Pile-T5 Large is an Encoder-Decoder model trained on the Pile using the T5x library. The model was trained for 2 million steps or roughly 2 trillion tokens using MLM-objective similar to the original T5 model.
The HF version of Pile-T5 Large borrows UMT5's model implementation as it uses scalable model implementation from T5x and uses LlamaTokenizer.
Model Details
- Developed by: EleutherAI
- Model type: Transformer-based Language Model
- Language: English
- Learn more: Blogpost. For details about the training dataset, see the Pile paper, and its data sheet.
- License: Apache 2.0
- Contact: to ask questions about this model, join the EleutherAI
Discord, and post them in
#release-discussion. Please read the existing GPT-NeoX-20B documentation before asking about the model on Discord. For general correspondence: contact@eleuther. ai.
| Hyperparameter | Value |
|---|---|
| nparameters | 783173632 |
| nencoder layers | 24 |
| ndecoder layers | 24 |
| dmodel | 2816 |
| demb | 1024 |
| nheads | 16 |
| dhead | 64 |
| nvocab | 32128 |
| Sequence Length | 512 |
Uses and limitations
Intended use
Pile-T5 was developed primarily for research purposes. It learns an inner representation of the English language that can be used to extract features useful for downstream tasks.
In addition to scientific uses, you may also further fine-tune and adapt Pile-T5 for deployment, as long as your use is in accordance with the Apache 2.0 license. This model works with the Transformers Library. If you decide to use pre-trained Pile-T5 as a basis for your fine-tuned model, please note that you need to conduct your own risk and bias assessment.
Out-of-scope use
Pile-T5 is not intended for deployment as-is. It is not a product and cannot be used for human-facing interactions without supervision.
Pile-T5 has not been fine-tuned for downstream tasks for which language models are commonly deployed, such as writing genre prose, or commercial chatbots. This means Pile-T5 will likely not respond to a given prompt the way products such as ChatGPT do. This is because, unlike Pile-T5, ChatGPT was fine-tuned using methods such as Reinforcement Learning from Human Feedback (RLHF) to better “understand” human instructions and dialogue.
This model is English-language only, and thus cannot be used for translation or generating text in other languages.
Limitations and biases
The core functionality of Pile-T5 is to take a string of text that has been partially replaced with mask tokens and predict a sequence of tokens that would replace those mask tokens. Remember that the statistically most likely sequence of tokens need not result in the most “accurate” text. Never rely on Pile-T5 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. Pile-T5 may produce socially unacceptable or undesirable text, even if the prompt itself does not include anything explicitly offensive.
We recommend curating the outputs of this model before presenting it to a human reader. Please inform your audience that you are using artificially generated text.
How to use
Pile-T5 can be loaded using the AutoModelForSeq2SeqLM functionality:
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tokenizer = AutoTokenizer.from_pretrained("EleutherAI/pile-t5-large")
model = AutoModelForSeq2SeqLM.from_pretrained("EleutherAI/pile-t5-large")
Training
Training dataset
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.
The Pile was deduplicated before being used to train Pile-T5.
Training procedure
Pile-T5 was trained with a batch size of approximately 1M tokens (2048 sequences of 512 tokens each), for a total of 2,000,000 steps. Pile-T5 was trained with the span-corruption objective.
Training checkpoints
Intermediate checkpoints for Pile-T5 are accessible within this repository. There are in total 200 checkpoints that are spaced 10,000 steps. For T5x-native checkpoints that can be used for finetuning with the T5x library, refer to here
The training loss (in tfevent format) and validation perplexity (in jsonl) can be found here.
Evaluations
Pile-T5 Large was evaluated on SuperGLUE, CodeXGLUE. A Flan-finetuned version was evaluated on Flan Held In tasks. Results can be seen in the blogpost
BibTeX
@misc{2024PileT5,
author = {Lintang Sutawika and Aran Komatsuzaki and Colin Raffel},
title = {Pile-T5},
year = {2024},
url = {https://blog.eleuther.ai/pile-t5/},
note = {Blog post},
}