README.md

metadata

language: en
tags:
  - azbert
  - pretraining
  - fill-mask
widget:
  - text: $f$ $($ $x$ [MASK] $y$ $)$
    example_title: mathy
  - text: $x$ [MASK] $x$ $equal$ $2$ $x$
    example_title: mathy
  - text: Proof by [MASK] that $n$ $fact$ $gt$ $3$ $n$ for $n$ $gt$ $6$
    example_title: mathy
  - text: Proof by induction that $n$ [MASK] $gt$ $3$ $n$ for $n$ $gt$ $6$
    example_title: mathy
  - text: The goal of life is [MASK].
    example_title: philosophical
license: mit

About

Here we share a pretrained BERT model that is aware of math tokens. The math tokens are treated specially and tokenized using pya0, which adds very limited new tokens for latex markup (total vocabulary is just 31,061).

This model is trained on 4 x 2 Tesla V100 with a total batch size of 64, using Math StackExchange data with 2.7 million sentence pairs trained for 7 epochs.

Usage

Download and try it out

pip install pya0==0.3.2
wget https://vault.cs.uwaterloo.ca/s/gqstFZmWHCLGXe3/download -O ckpt.tar.gz
mkdir -p ckpt
tar xzf ckpt.tar.gz -C ckpt --strip-components=1
python test.py --test_file test.txt

Test file format

Modify the test examples in test.txt to play with it.

The test file is tab-separated, the first column is additional positions you want to mask for the right-side sentence (useful for masking tokens in math markups). A zero means no additional mask positions.

Example output

Upload to huggingface

This repo is hosted on Github, and only mirrored at huggingface.

To upload to huggingface, use the upload2hgf.sh script. Before runnig this script, be sure to check:

• check points for model and tokenizer are created under ./ckpt folder
• model contains all the files needed: config.json and pytorch_model.bin
• tokenizer contains all the files needed: added_tokens.json, special_tokens_map.json, tokenizer_config.json, vocab.txt and tokenizer.json
• no tokenizer_file field in tokenizer_config.json (sometimes it is located locally at ~/.cache)
• git-lfs is installed
• having git-remote named hgf reference to https://huggingface.co/castorini/azbert-base