Hugging Face's logo

ckip-joint
/
bloom-1b1-zh

Text Generation
Transformers
PyTorch
Chinese
bloom
feature-extraction
text-generation-inference
Inference Endpoints
Model card Files Community
6
bloom-1b1-zh / README.md
pochunhsu's picture
pochunhsu
Update README.md
1b0edc1
preview code
|
raw
history blame
6.86 kB
metadata
license: bigscience-bloom-rail-1.0
language:
  - en
  - zht
pipeline_tag: text-generation

BLOOM-zh

Open-access Multilingual Language Model based on BLOOM

Model Card

Version 1.0 / 20.Feb.2023

This model is a joint collaboration between CKIP lab at Acedemia Sinica (website), MediaTek Research (website), and National Academy for Educational Research (website).

Table of Contents

  1. Model Details
  2. Uses
  3. Training Data
  4. Risks and Limitations
  5. Evaluation
  6. Recommendations
  7. Glossary and Calculations
  8. More Information
  9. Model Card Authors

Model Details

BLOOM-zh is a modification from BLOOMZ. BLOOM-zh is trained extendedly on larger amounts of Traditional Chinese text data while it still maintains its pretrained English ability.

Basics

This section provides information for anyone who wants to know about the model.

Click to expand

Developed by: MediaTek Research

Model Type: Transformer-based Language Model

Version: 1.0.0

Languages: Multiple; see training data

License: MEDIATEK RESEARCH License (link) and RAIL License v1.0 (link)

Release Date Estimate: Wednesday, 22.February.2023

Send Questions to: info@mtkresearch.com

Cite as: MediaTek Research, MediaTek Research Open-access Multilingual Language Model based on BLOOM. International, February 2023.

Organizations of contributors:

  • MediaTek Research
  • Academia Sinica
  • National Academy for Educational Research

Technical Specifications

This section provides information for people who work on model development.

Click to expand

Model Architecture: 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

  • 1,065,314,304 parameters:

    • 385,351,680 embedding parameters

    • 24 layers, 16 attention heads

    • Hidden layers are 1536-dimensional

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

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

Compute infrastructure:

Training

Details are provided in the paper.

  • Dates: Feb. 2023

Tokenization

The BLOOM tokenizer (link) is 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.

Environmental Impact

Click to expand

Please refer to Model card.

 

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 provides information for anyone considering using the model or who is affected by the model.

Click to expand

Please refer to Model card.

 

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.

Click to expand

We trained the 1B1 parameter model on a total of 6 Billion tokens mainly crawled from the internet and provided from National Academy for Educational Research. 75% of the training data is Traditional Chinese, 25% is English. Details are provided in the paper.

 

Risks and Limitations

This section identifies foreseeable harms and misunderstandings.

Click to expand

Please refer to Model card.

 

Factors

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

  • The model is trained on Traditional Chinese and English. However, the pretrained weights capture more than 40 different languages.

  • The model is trained on web crawled data, news articles, novels, knowledge sources (encyclopedia, education sector) and instructions

 

Recommendations

This section provides information on warnings and potential mitigations.

Click to expand

Please refer to Model card.

 

Model Card Authors

Ordered roughly chronologically and by amount of time spent.

Philipp Ennen, Po-Chun Hsu, Chan-Jan Hsu, Chang-Le Liu, Yin-Hsiang Liao, Chin-Tung Lin, Jezabel Rodriguez Garcia, Federica Freddi, Da-Shan Shiu, Wei-Yun Ma