mathy-vicuna-13B-FFT

About

This repository is a boilerplate to push a mask-filling model to the HuggingFace Model Hub.

Upload to huggingface

Download your tokenizer, model checkpoints, and optionally the training logs (events.out.*) to the ./ckpt directory (do not include any large files except pytorch_model.bin and log files events.out.*).

Optionally, test model using the MLM task:

pip install pya0 # for math token preprocessing
# testing local checkpoints:
python test.py ./ckpt/math-tokenizer ./ckpt/2-2-0/encoder.ckpt
# testing Model Hub checkpoints:
python test.py approach0/coco-mae-220 approach0/coco-mae-220

Note
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.

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

• git-lfs is installed
• having git-remote named hgf reference to https://huggingface.co/your/repo
• 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)