Zelda Rose

A straightforward trainer for transformer-based models.

Installation

Simply install with pipx

pipx install zeldarose

Train MLM models

Here is a short example of training first a tokenizer, then a transformer MLM model:

TOKENIZERS_PARALLELISM=true zeldarose tokenizer --vocab-size 4096 --out-path local/tokenizer  --model-name "my-muppet" tests/fixtures/raw.txt
zeldarose 
transformer --tokenizer local/tokenizer --pretrained-model flaubert/flaubert_small_cased --out-dir local/muppet --val-text tests/fixtures/raw.txt tests/fixtures/raw.txt

The .txt files are meant to be raw text files, with one sample (e.g. sentence) per line.

There are other parameters (see zeldarose transformer --help for a comprehensive list), the one you are probably mostly interested in is --config, giving the path to a training config (for which we have examples/).

The parameters --pretrained-models, --tokenizer and --model-config are all fed directly to Huggingface's transformers and can be pretrained models names or local path.

Distributed training

This is somewhat tricky, you have several options

• If you are running in a SLURM cluster use --strategy ddp and invoke via srun

  • You might want to preprocess your data first outside of the main compute allocation. The --profile option might be abused for that purpose, since it won't run a full training, but will run any data preprocessing you ask for. It might also be beneficial at this step to load a placeholder model such as RoBERTa-minuscule to avoid runnin out of memory, since the only thing that matter for this preprocessing is the tokenizer.

• Otherwise you have two options

  • Run with --strategy ddp_spawn, which uses multiprocessing.spawn to start the process swarm (tested, but possibly slower and more limited, see pytorch-lightning doc)
  • Run with --strategy ddp and start with torch.distributed.launch with --use_env and --no_python (untested)

Other hints

• Data management relies on 🤗 datasets and use their cache management system. To run in a clear environment, you might have to check the cache directory pointed to by theHF_DATASETS_CACHE environment variable.

Inspirations

• https://github.com/shoarora/lmtuners
• https://github.com/huggingface/transformers/blob/243e687be6cd701722cce050005a2181e78a08a8/examples/run_language_modeling.py

Citation

@inproceedings{grobol:hal-04262806,
    TITLE = {{Zelda Rose: a tool for hassle-free training of transformer models}},
    AUTHOR = {Grobol, Lo{\"i}c},
    URL = {https://hal.science/hal-04262806},
    BOOKTITLE = {{3rd Workshop for Natural Language Processing Open Source Software (NLP-OSS)}},
    ADDRESS = {Singapore, Indonesia},
    YEAR = {2023},
    MONTH = Dec,
    PDF = {https://hal.science/hal-04262806/file/Zeldarose_OSS_EMNLP23.pdf},
    HAL_ID = {hal-04262806},
    HAL_VERSION = {v1},
}