fedLM

A federated LM

Electra

Setup

Clone from original source

git clone [email protected]:google-research/electra.git

Then use docker-compose files in test folder to run the preprocessing and other necessary steps. First move them into the electra folder, then build and run.

Build the docker container:

docker-compose -f docker-compose.yaml build

and run it interactively

docker-compose run -u $(id -u):$(id -g) --rm electra bash

Pretraining

Create shards and put everything into the TF format.

python3 build_pretraining_dataset.py --corpus-dir ../data/${lang} --vocab-file ../data/vocab.${lang}.txt --output-dir ./data/ --max-seq-length 128 --num-processes 15 --blanks-separate-docs True --do-lower-case

Then run the pre-training:

python3 run_pretraining.py --data-dir data/ --model-name electra_small_$lang --hparams '{"debug": false, "do_train": true, "do_eval": false, "vocab_size": 31000, "vocab_file": "vocab.$lang.txt"}'

Robin Forgets

• kb-labb-1 Danish
• kb-labb-2 Norwegian
• kb-labb-3 Swedish

Checkpoints @

• 50% - 500k
• 25% - 250k
• 12.5% - 125k
• 6.25% - 62.5k

Time Estimates

Strongly depends on GPU (yay RTX 3090).

RTX 3090 : 13.5k / h RTX 2080 : 7.8k / h RTX 2060 : ?.?k / h