WARNING: The checkpoints on this repo are not fully trained model. Evaluations of intermediary checkpoints and the final model will be added when conducted (see below).

BLOOM LM BigScience Large Open-science Open-access Multilingual Language Model Model Card

Version 1.3 / 11.July.2022 - Available intermediary checkpoints - global steps:

• 5000 , 10000 , 15000 , 20000 , 25000 , 30000 , 35000 , 40000 , 45000 , 50000 , 55000 , 60000 , 65000 , 70000 , 75000 , 80000 , 85000 , 90000 , 91000 , 92000 , 93000

You can check the available checkpoints by clicking on the branches section of the repo

How to load a specific version

We use git tags to load a model in a specific version (eg. global_step5000 ):

Table of Contents

Model Details

Uses

Training Data

Risks and Limitations

Evaluation

Recommendations

Glossary and Calculations

More Information

Model Card Authors

Model Details

BLOOM is a type of language model, which is a probability distribution over sequences of words. Specifically, BLOOM is a Large Language Model (LLM), meaning that it is trained on vast amounts of text data using industrial-scale computational resources. As such, the model is able to capture the statistical tendencies of words, phrases, sentences, and larger spans of text that it is exposed to in the training data.

Basics

This section provides information about the model type, version, license, funders, release date, developers, and contact information. It is useful for anyone who wants to reference the model.

Click to expand

Developed by: BigScience ( website )

All collaborators are either volunteers or have an agreement with their employer. (Further breakdown of participants forthcoming.)

Model Type: Transformer-based Language Model

Version: 1.0.0

Languages: Multiple; see training data

License: RAIL License v1.0 ( link )

Release Date Estimate: Monday, 11.July.2022

Send Questions to: bigscience-contact@googlegroups.com

Cite as: BigScience, BigScience Language Open-science Open-access Multilingual (BLOOM) Language Model . International, May 2021-May 2022

Funded by:

• The French government.

• Hugging Face ( website ).

• Organizations of contributors. (Further breakdown of organizations forthcoming.)

Technical Specifications

This section includes details about the model objective and architecture, and the compute infrastructure. It is useful for people interested in model development.

Click to expand

Please see the BLOOM training README for full details on replicating training.

Model Architecture and Objective

• Modified from Megatron-LM GPT2 (see paper , BLOOM Megatron code ):

• Decoder-only architecture

• Layer normalization applied to word embeddings layer ( StableEmbedding ; see code , paper )

• ALiBI positional encodings (see paper ), with GeLU activation functions

• 176 billion parameters:

  • 70 layers, 112 attention heads

  • Hidden layers are 14336-dimensional

  • Sequence length of 2048 tokens used (see BLOOM tokenizer , tokenizer description )

Objective Function: Cross Entropy with mean reduction (see API documentation ).

Compute infrastructure

Jean Zay Public Supercomputer, provided by the French government (see announcement ).

Hardware

• 384 A100 80GB GPUs (48 nodes)

• Additional 32 A100 80GB GPUs (4 nodes) in reserve

• 8 GPUs per node Using NVLink 4 inter-gpu connects, 4 OmniPath links

• CPU: AMD

• CPU memory: 512GB per node

• GPU memory: 640GB per node

• Inter-node connect: Omni-Path Architecture (OPA)

• NCCL-communications network: a fully dedicated subnet

• Disc IO network: shared network with other types of nodes

Software

• Megatron-DeepSpeed ( Github link )

• DeepSpeed ( Github link )

• PyTorch (pytorch-1.11 w/ CUDA-11.5; see Github link )

• apex ( Github link )

Training

This section provides information about the training data, the speed and size of training elements, and the environmental impact of training. It is useful for people who want to learn more about the model inputs and training footprint.

Click to expand

Training Data

This section provides a high-level overview of the training data. It is relevant for anyone who wants to know the basics of what the model is learning.

Details for each dataset are provided in individual Data Cards .

Training data includes:

• 45 natural languages

• 12 programming languages

• In 1.5TB of pre-processed text, converted into 350B unique tokens (see the tokenizer section for more.)

Languages

The pie chart shows the distribution of languages in training data.

The following tables shows the further distribution of Niger-Congo & Indic languages and programming languages in the training data.

Distribution of Niger Congo and Indic languages.

Niger Congo	Percentage	Indic	Percentage
Chi Tumbuka	0.00002	Assamese	0.01
Kikuyu	0.00004	Odia	0.04
Bambara	0.00004	Gujarati	0.04
Akan	0.00007	Marathi	0.05
Xitsonga	0.00007	Punjabi	0.05
Sesotho	0.00007	Kannada	0.06
Chi Chewa	0.0001	Nepali	0.07
Setswana	0.0002	Telugu	0.09
Northern Sotho	0.0002	Malayalam	0.10
Fon	0.0002	Urdu	0.10
Kirundi	0.0003	Tamil	0.20
Wolof	0.0004	Bengali	0.50
Kuganda	0.0004	Hindi	0.70
Chi Shona	0.001
Isi Zulu	0.001
Igbo	0.001
Xhosa	0.001
Kinyarwanda	0.003
Yoruba	0.006
Swahili	0.02

Distribution of programming languages.

Extension	Language	Number of files
java	Java	5,407,724
php	PHP	4,942,186
cpp	C++	2,503,930
py	Python	2,435,072
js	JavaScript	1,905,518
cs	C#	1,577,347
rb	Ruby	6,78,413
cc	C++	443,054
hpp	C++	391,048
lua	Lua	352,317
go	GO	227,763
ts	TypeScript	195,254
C	C	134,537
scala	Scala	92,052
hh	C++	67,161
H	C++	55,899
tsx	TypeScript	33,107
rs	Rust	29,693
phpt	PHP	9,702
c++	C++	1,342
h++	C++	791
php3	PHP	540
phps	PHP	270
php5	PHP	166
php4	PHP	29

Preprocessing

Tokenization: The BLOOM tokenizer ( link ), a learned subword tokenizer trained using:

• A byte-level Byte Pair Encoding (BPE) algorithm

• A simple pre-tokenization rule, no normalization

• A vocabulary size of 250,680

It was trained on a subset of a preliminary version of the corpus using alpha-weighting per language.

Speeds, Sizes, Times

Training logs: Tensorboard link

• Dates:

  • Started 11th March, 2022 11:42am PST

  • Estimated end: 5th July, 2022

• Checkpoint size:

  • Bf16 weights: 329GB

  • Full checkpoint with optimizer states: 2.3TB

• Training throughput: About 150 TFLOP per GPU per second

• Number of epochs: 1

• Estimated cost of training: Equivalent of $2-5M in cloud computing (including preliminary experiments)

• Server training location: Île-de-France, France

Environmental Impact

The training supercomputer, Jean Zay ( website ), uses mostly nuclear energy. The heat generated by it is reused for heating campus housing.

Estimated carbon emissions: (Forthcoming.)

Estimated electricity usage: (Forthcoming.)

Uses

This section addresses questions around how the model is intended to be used, discusses the foreseeable users of the model (including those affected by the model), and describes uses that are considered out of scope or misuse of the model. It is useful for anyone considering using the model or who is affected by the model.

Click to expand

Intended Use

This model is being created in order to enable public research on large language models (LLMs). LLMs are intended to be used for language generation or as a pretrained base model that can be further fine-tuned for specific tasks. Use cases below are not exhaustive.

Direct Use

• Text generation

• Exploring characteristics of language generated by a language model

  • Examples: Cloze tests, counterfactuals, generations with reframings

Downstream Use

• Tasks that leverage language models include: Information Extraction, Question Answering, Summarization

Misuse and Out-of-scope Use

This section addresses what users ought not do with the model.

See the BLOOM License , Attachment A, for detailed usage restrictions. The below list is non-exhaustive, but lists some easily foreseeable problematic use cases.

Out-of-scope Uses

Using the model in high-stakes settings is out of scope for this model. The model is not designed for critical decisions nor uses with any material consequences on an individual's livelihood or wellbeing. The model outputs content that appears factual but is not correct.

Out-of-scope Uses Include:

• Usage in biomedical domains, political and legal domains, or finance domains

• Usage for evaluating or scoring individuals, such as for employment, education, or credit

• Applying the model for critical automatic decisions, generating factual content, creating reliable summaries, or generating predictions that must be correct

Misuse

Intentionally using the model for harm, violating human rights , or other kinds of malicious activities, is a misuse of this model. This includes:

• Spam generation

• Disinformation and influence operations

• Disparagement and defamation

• Harassment and abuse

• Deception

• Unconsented impersonation and imitation

• Unconsented surveillance

• Generating content without attribution to the model, as specified in the RAIL License, Use Restrictions

Intended Users

Direct Users

• General Public

• Researchers

• Students

• Educators

• Engineers/developers

• Non-commercial entities

• Community advocates, including human and civil rights groups

Indirect Users

• Users of derivatives created by Direct Users, such as those using software with an intended use

• Users of Derivatives of the Model, as described in the License

Others Affected (Parties Prenantes)

• People and groups referred to by the LLM

• People and groups exposed to outputs of, or decisions based on, the LLM

• People and groups whose original work is included in the LLM

Risks and Limitations

This section identifies foreseeable harms and misunderstandings.

Click to expand

Model may:

• Overrepresent some viewpoints and underrepresent others

• Contain stereotypes

• Contain personal information

• Generate:

  • Hateful, abusive, or violent language

  • Discriminatory or prejudicial language

  • Content that may not be appropriate for all settings, including sexual content

• Make errors, including producing incorrect information as if it were factual

• Generate irrelevant or repetitive outputs

Evaluation

This section describes the evaluation protocols and provides the results.

Click to expand

Metrics

This section describes the different ways performance is calculated and why.

Includes:

Metric	Why chosen
Perplexity	Standard metric for quantifying model improvements during training
Cross Entropy Loss	Standard objective for language models.

And multiple different metrics for specific tasks. (More evaluation metrics forthcoming upon completion of evaluation protocol.)

Factors

This section lists some different aspects of what BLOOM models. Its focus is on those aspects that are likely to give rise to high variance in model behavior.

• Language, such as English or Yoruba

• Domain, such as newswire or stories

• Demographic characteristics, such as gender or nationality

Results

Results are based on the Factors and Metrics .

Train-time Evaluation:

As of 25.May.2022, 15:00 PST:

• Training Loss: 2.0

• Validation Loss: 2.2

• Perplexity: 8.9

(More evaluation scores forthcoming.)

Recommendations

This section provides information on warnings and potential mitigations.

Click to expand

• Indirect users should be made aware when the content they're working with is created by the LLM.

• Users should be aware of Risks and Limitations , and include an appropriate age disclaimer or blocking interface as necessary.

• Models trained or finetuned downstream of BLOOM LM should include an updated Model Card.

• Users of the model should provide mechanisms for those affected to provide feedback, such as an email address for comments.

Glossary and Calculations

This section defines common terms and how metrics are calculated.

Click to expand

• Loss: A calculation of the difference between what the model has learned and what the data shows ("groundtruth"). The lower the loss, the better. The training process aims to minimize the loss.

• Perplexity: This is based on what the model estimates the probability of new data is. The lower the perplexity, the better. If the model is 100% correct at predicting the next token it will see, then the perplexity is 1. Mathematically this is calculated using entropy.

• High-stakes settings: Such as those identified as "high-risk AI systems" and "unacceptable risk AI systems" in the European Union's proposed Artificial Intelligence (AI) Act .

• Critical decisions: Such as those defined in the United States' proposed Algorithmic Accountability Act .

• Human rights: Includes those rights defined in the Universal Declaration of Human Rights .

• Personal Data and Personal Information: Personal data and information is defined in multiple data protection regulations, such as " personal data " in the European Union's General Data Protection Regulation ; and "personal information" in the Republic of South Africa's Protection of Personal Information Act , The People's Republic of China's Personal information protection law .

• Sensitive characteristics: This includes specifically protected categories in human rights (see UHDR, Article 2 ) and personal information regulation (see GDPR, Article 9; Protection of Personal Information Act, Chapter 1 )

• Deception: Doing something to intentionally mislead individuals to believe something that is false, such as by creating deadbots or chatbots on social media posing as real people, or generating text documents without making consumers aware that the text is machine generated.

More Information

This section provides links to writing on dataset creation, technical specifications, lessons learned, and initial results.

Click to expand

Dataset Creation

Blog post detailing the design choices during the dataset creation: https://bigscience.huggingface.co/blog/building-a-tb-scale-multilingual-dataset-for-language-modeling

Technical Specifications

Blog post summarizing how the architecture, size, shape, and pre-training duration where selected: https://bigscience.huggingface.co/blog/what-language-model-to-train-if-you-have-two-million-gpu-hours

More details on the architecture/optimizer: https://github.com/bigscience-workshop/bigscience/tree/master/train/tr11-176B-ml

Blog post on the hardware/engineering side: https://bigscience.huggingface.co/blog/which-hardware-to-train-a-176b-parameters-model

Details on the distributed setup used for the training: https://github.com/bigscience-workshop/bigscience/tree/master/train/tr11-176B-ml

Tensorboard updated during the training: https://huggingface.co/bigscience/tr11-176B-ml-logs/tensorboard#scalars&tagFilter=loss

Lessons

Insights on how to approach training, negative results: https://github.com/bigscience-workshop/bigscience/blob/master/train/lessons-learned.md

Details on the obstacles overcome during the preparation on the engineering side (instabilities, optimization of training throughput, so many technical tricks and questions): https://github.com/bigscience-workshop/bigscience/blob/master/train/tr11-176B-ml/chronicles.md

Initial Results

Initial prompting experiments using interim checkpoints: https://huggingface.co/spaces/bigscience/bloom-book

Model Card Authors

Ordered roughly chronologically and by amount of time spent.

Margaret Mitchell, Giada Pistilli, Yacine Jernite, Ezinwanne Ozoani, Marissa Gerchick, Nazneen Rajani, Sasha Luccioni, Irene Solaiman, Maraim Masoud, Somaieh Nikpoor, Carlos Muñoz Ferrandis, Stas Bekman, Christopher Akiki, Danish Contractor, David Lansky, Angelina McMillan-Major, Tristan Thrush, Suzana Ilić, Gérard Dupont, Shayne Longpre, Manan Dey, Stella Biderman, Douwe Kiela, Emi Baylor, Teven Le Scao, Aaron Gokaslan, Julien Launay