Edit model card
BigScience Logo

BigScience Large Open-science Open-access Multilingual Language Model
Version 1.3 / 6 July 2022

Current Checkpoint: Training Iteration 95000

Total seen tokens: 366B


Model Details

BLOOM is an autoregressive Large Language Model (LLM), trained to continue text from a prompt on vast amounts of text data using industrial-scale computational resources. As such, it is able to output coherent text in 46 languages and 13 programming languages that is hardly distinguishable from text written by humans. BLOOM can also be instructed to perform text tasks it hasn't been explicitly trained for, by casting them as text generation tasks.

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

Checkpoints format: transformers (Megatron-DeepSpeed format available here)

Version: 1.0.0

Languages: Multiple; see training data

License: RAIL License v1.0 (link / article and FAQ)

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:

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


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, and the sizes of each of their contributions to the aggregated training data are presented in an Interactive Corpus Map.

Training data includes:

  • 46 natural languages

  • 13 programming languages

  • In 1.6TB 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.

pie chart showing 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
Lingala 0.0002 Malayalam 0.10
Northern Sotho 0.0002 Urdu 0.10
Fon 0.0002 Tamil 0.20
Kirundi 0.0003 Bengali 0.50
Wolof 0.0004 Hindi 0.70
Luganda 0.0004
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

How to use

This model can be easily used and deployed using HuggingFace's ecosystem. This needs transformers and accelerate installed. The model can be downloaded as follows:

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 may not be 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

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

  • Induce users into attributing human traits to it, such as sentience or consciousness


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 BLOOM models. Its focus is on 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.

Zero-shot evaluations:

WARNING: These are intermediate results

See this repository for JSON files: https://github.com/bigscience-workshop/evaluation-results

Task Language Metric BLOOM-176B OPT-175B*
arc_challenge eng acc ↑ 0.411 0.412
arc_easy eng acc ↑ 0.726 0.751
axb (Median of 10 prompts) eng acc ↑ 0.575 0.532
axg (Median of 10 prompts) eng acc ↑ 0.525 0.548
boolq (Median of 11 prompts) eng acc ↑ 0.635 0.622
cb (Median of 15 prompts) eng acc ↑ 0.339 0.411
cola (Median of 5 prompts) eng acc ↑ 0.39 0.444
copa (Median of 9 prompts) eng acc ↑ 0.56 0.55
crows_pairs_english (Median of 6 prompts) eng acc ↑ 0.5 0.502
crows_pairs_french (Median of 7 prompts) fra acc ↑ 0.506 0.499
diabla (Median of 2 prompts) eng acc ↑ 0.295 0.289
gsarti/flores_101_afr afr byte_perplexity ↓ 4.254 3.381
gsarti/flores_101_amh amh byte_perplexity ↓ 3.717 3.87
gsarti/flores_101_ara ara byte_perplexity ↓ 1.705 2.42
gsarti/flores_101_asm asm byte_perplexity ↓ 6.577 3.028
gsarti/flores_101_ast ast byte_perplexity ↓ 2.856 4.737
gsarti/flores_101_azj azj byte_perplexity ↓ 4.807 4.767
gsarti/flores_101_bel bel byte_perplexity ↓ 2.731 2.557
gsarti/flores_101_ben ben byte_perplexity ↓ 5.993 2.243
gsarti/flores_101_bos bos byte_perplexity ↓ 3.594 2.668
gsarti/flores_101_bul bul byte_perplexity ↓ 2.159 2.099
gsarti/flores_101_cat cat byte_perplexity ↓ 2.168 2.837
gsarti/flores_101_ceb ceb byte_perplexity ↓ 5.287 3.636
gsarti/flores_101_ces ces byte_perplexity ↓ 3.452 2.749
gsarti/flores_101_ckb ckb byte_perplexity ↓ 3.705 4.688
gsarti/flores_101_cym cym byte_perplexity ↓ 7.089 5.075
gsarti/flores_101_dan dan byte_perplexity ↓ 3.43 2.492
gsarti/flores_101_deu deu byte_perplexity ↓ 2.338 2.099
gsarti/flores_101_ell ell byte_perplexity ↓ 1.96 1.811
gsarti/flores_101_eng eng byte_perplexity ↓ 1.882 1.9
gsarti/flores_101_est est byte_perplexity ↓ 5.774 3.533
gsarti/flores_101_fas fas byte_perplexity ↓ 2.431 2.444
gsarti/flores_101_fin fin byte_perplexity ↓ 4.304 2.601
gsarti/flores_101_fra fra byte_perplexity ↓ 1.937 1.984
gsarti/flores_101_ful ful byte_perplexity ↓ 9.74 11.84
gsarti/flores_101_gle gle byte_perplexity ↓ 6.035 3.914
gsarti/flores_101_glg glg byte_perplexity ↓ 2.365 3.015
gsarti/flores_101_guj guj byte_perplexity ↓ 5.707 2.438
gsarti/flores_101_hau hau byte_perplexity ↓ 8.855 5.283
gsarti/flores_101_heb heb byte_perplexity ↓ 2.921 2.903
gsarti/flores_101_hin hin byte_perplexity ↓ 5.452 1.86
gsarti/flores_101_hrv hrv byte_perplexity ↓ 3.706 2.715
gsarti/flores_101_hun hun byte_perplexity ↓ 4.059 2.865
gsarti/flores_101_hye hye byte_perplexity ↓ 3.127 3.411
gsarti/flores_101_ibo ibo byte_perplexity ↓ 3.95 8.008
gsarti/flores_101_ind ind byte_perplexity ↓ 1.976 2.632
gsarti/flores_101_isl isl byte_perplexity ↓ 5.501 4.701
gsarti/flores_101_ita ita byte_perplexity ↓ 2.314 2.104
gsarti/flores_101_jav jav byte_perplexity ↓ 4.942 8.16
gsarti/flores_101_jpn jpn byte_perplexity ↓ 2.259 2.198
gsarti/flores_101_kam kam byte_perplexity ↓ 9.743 10.981
gsarti/flores_101_kan kan byte_perplexity ↓ 6.234 2.373
gsarti/flores_101_kat kat byte_perplexity ↓ 2.051 2.466
gsarti/flores_101_kaz kaz byte_perplexity ↓ 3.039 4.376
gsarti/flores_101_kea kea byte_perplexity ↓ 7.147 9.632
gsarti/flores_101_khm khm byte_perplexity ↓ 3.367 2.646
gsarti/flores_101_kir kir byte_perplexity ↓ 3.241 4.522
gsarti/flores_101_kor kor byte_perplexity ↓ 2.902 3.376
gsarti/flores_101_lao lao byte_perplexity ↓ 2.331 3.106
gsarti/flores_101_lav lav byte_perplexity ↓ 5.224 4.811
gsarti/flores_101_lin lin byte_perplexity ↓ 4.847 8.871
gsarti/flores_101_lit lit byte_perplexity ↓ 4.543 5.183
gsarti/flores_101_ltz ltz byte_perplexity ↓ 5.591 7.158
gsarti/flores_101_lug lug byte_perplexity ↓ 5.43 7.399
gsarti/flores_101_luo luo byte_perplexity ↓ 12.031 11.951
gsarti/flores_101_mal mal byte_perplexity ↓ 4.794 2.054
gsarti/flores_101_mar mar byte_perplexity ↓ 6.857 2.274
gsarti/flores_101_mkd mkd byte_perplexity ↓ 2.335 2.538
gsarti/flores_101_mlt mlt byte_perplexity ↓ 9.041 5.996
gsarti/flores_101_mon mon byte_perplexity ↓ 3.095 4.519
gsarti/flores_101_mri mri byte_perplexity ↓ 5.266 4.438
gsarti/flores_101_msa msa byte_perplexity ↓ 2.222 2.935
gsarti/flores_101_mya mya byte_perplexity ↓ 2.523 2.413
gsarti/flores_101_nld nld byte_perplexity ↓ 2.799 2.293
gsarti/flores_101_nob nob byte_perplexity ↓ 3.629 2.593
gsarti/flores_101_npi npi byte_perplexity ↓ 6.666 2.499
gsarti/flores_101_nso nso byte_perplexity ↓ 5.015 8.485
gsarti/flores_101_nya nya byte_perplexity ↓ 4.938 7.548
gsarti/flores_101_oci oci byte_perplexity ↓ 3.607 4.936
gsarti/flores_101_orm orm byte_perplexity ↓ 11.316 7.145
gsarti/flores_101_ory ory byte_perplexity ↓ 5.982 2.668
gsarti/flores_101_pan pan byte_perplexity ↓ 4.772 2.782
gsarti/flores_101_pol pol byte_perplexity ↓ 3.012 2.432
gsarti/flores_101_por por byte_perplexity ↓ 1.841 2.178
gsarti/flores_101_pus pus byte_perplexity ↓ 4.624 4.785
gsarti/flores_101_ron ron byte_perplexity ↓ 3.05 2.197
gsarti/flores_101_rus rus byte_perplexity ↓ 1.708 1.689
gsarti/flores_101_slk slk byte_perplexity ↓ 4.038 3.419
gsarti/flores_101_slv slv byte_perplexity ↓ 4.141 3.582
gsarti/flores_101_sna sna byte_perplexity ↓ 4.711 5.588
gsarti/flores_101_snd snd byte_perplexity ↓ 4.206 5.667
gsarti/flores_101_som som byte_perplexity ↓ 9.154 4.788
gsarti/flores_101_spa spa byte_perplexity ↓ 1.796 2.098
gsarti/flores_101_srp srp byte_perplexity ↓ 2.241 2.688
gsarti/flores_101_swe swe byte_perplexity ↓ 3.345 2.468
gsarti/flores_101_swh swh byte_perplexity ↓ 2.684 4.473
gsarti/flores_101_tam tam byte_perplexity ↓ 5.165 2.024
gsarti/flores_101_tel tel byte_perplexity ↓ 6.81 2.407
gsarti/flores_101_tgk tgk byte_perplexity ↓ 3.785 4.899
gsarti/flores_101_tgl tgl byte_perplexity ↓ 3.75 2.738
gsarti/flores_101_tha tha byte_perplexity ↓ 2.104 2.035
gsarti/flores_101_tur tur byte_perplexity ↓ 3.318 2.622
gsarti/flores_101_ukr ukr byte_perplexity ↓ 2.089 1.93
gsarti/flores_101_umb umb byte_perplexity ↓ 11.766 11.64
gsarti/flores_101_urd urd byte_perplexity ↓ 1.779 2.982
gsarti/flores_101_uzb uzb byte_perplexity ↓ 8.5 13.209
gsarti/flores_101_vie vie byte_perplexity ↓ 1.659 2.229
gsarti/flores_101_wol wol byte_perplexity ↓ 6.142 13.945
gsarti/flores_101_xho xho byte_perplexity ↓ 4.69 8.42
gsarti/flores_101_yor yor byte_perplexity ↓ 4.361 7.636
gsarti/flores_101_zho_simpl zho_simpl byte_perplexity ↓ 2.118 5.113
gsarti/flores_101_zho_trad zho_trad byte_perplexity ↓ 2.274 5.67
gsarti/flores_101_zul zul byte_perplexity ↓ 6.017 7.341
headqa esp acc ↑ 0.346 0.244
hellaswag eng acc ↑ 0.535 0.592
lambada_mt_de deu acc ↑ 0.329 0.358
lambada_mt_en eng acc ↑ 0.672 0.747
lambada_mt_es esp acc ↑ 0.476 0.397
lambada_mt_it ita acc ↑ 0.406 0.409
logiqa eng acc ↑ 0.235 0.244
mathqa eng acc ↑ 0.277 0.268
mc_taco eng em ↑ 0.131 0.124
mnli (Median of 15 prompts) eng acc ↑ 0.355 0.36
mnli_mismatched (Median of 15 prompts) eng acc ↑ 0.355 0.36
mrpc eng acc ↑ 0.387 0.446
multirc (Median of 11 prompts) eng acc ↑ 0.571 0.599
openbookqa eng acc ↑ 0.312 0.322
piqa eng acc ↑ 0.781 0.791
prost eng acc ↑ 0.298 0.299
pubmedqa eng acc ↑ 0.741 0.709
qnli eng acc ↑ 0.517 0.554
qqp (Median of 7 prompts) eng acc ↑ 0.588 0.395
race eng acc ↑ 0.39 0.402
rte (Median of 6 prompts) eng acc ↑ 0.52 0.495
sciq eng acc ↑ 0.936 0.948
sst (Median of 6 prompts) eng acc ↑ 0.604 0.647
triviaqa eng acc ↑ 0.183 0.342
tydiqa_primary (Median of 16 prompts) eng acc ↑ 0.281 0.148
webqs eng acc ↑ 0.062 0.159
wic (Median of 11 prompts) eng acc ↑ 0.506 0.498
winogrande eng acc ↑ 0.71 0.736
wnli (Median of 6 prompts) eng acc ↑ 0.57 0.563
wsc (Median of 11 prompts) eng acc ↑ 0.519 0.413
humaneval python pass@1 ↑ 0.155 0.0
humaneval python pass@10 ↑ 0.322 0.0
humaneval python pass@100 ↑ 0.555 0.003

Train-time Evaluation:

Final checkpoint after 95K steps:

  • Training Loss: 1.939

  • Validation Loss: 2.061

  • Perplexity: 7.045

For more see: https://huggingface.co/bigscience/tr11-176B-ml-logs


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

More Information

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

Click to expand

Intermediate checkpoints

For academic (or any) usage, we published the intermediate checkpoints, corresponding to the model state at each 5000 steps. Please follow this link to get these checkpoints.

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

Original checkpoints

The checkpoints in this repo correspond to the HuggingFace Transformers format. If you want to use our fork of Megatron-DeepSpeed that the model was trained with, you'd want to use this repo instead.


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, Niklas Muennighoff

Downloads last month
47,227
Hosted inference API
Text Generation
Groups
Groups
Examples
Examples
sampling
greedy
BLOOM prompting tips

Switch to "greedy" for more accurate completion e.g. math/history/translations (but which may be repetitive/less inventive)

Please login or register to try BLOOM 🌸
This model can be loaded on the Inference API on-demand.

Spaces using bigscience/bloom

Evaluation results