model-garden-lms/bert-base-token-dropping-finewebs-801k

FineWeb-LMs: Token Dropping BERT

This repository presents a Token Dropping BERT model that was pretrained on the 10BT subsets of FineWeb and FineWeb-Edu.

Pretraining Details

The released BERT model is part of my TensorFlow Model Garden LMs project.

The pretraining was done on a v3-32 TPU VM Pod, provided by the amazing TRC program. Detailed cheatsheets are available:

• TPU VM Setup
• Pretraining a Token Dropping BERT Model with TensorFlow Model Garden Library

tl;dr: The model was pretrained for 1M steps with a global batch size of 512, a sequence length of 512 using a vocab size of 64k.

Checkpoint Evaluation with ScandEval

We evaluate the last 5 checkpoints (1M, 951k, 901k, 851k and 851k) with a recent version of ScandEval to check their performance and also compare it with popular encoder-only models such as BERT, RoBERTa or ELECTRA:

Model ID	Avg. Score	CoNLL-En	SST5	ScaLA-En	SQuAD
model-garden-lms/bert-base-token-dropping-finewebs-1m	67.66	88.68 ± 0.76 / 88.47 ± 0.62	57.4 ± 1.7 / 59.61 ± 1.6	52.72 ± 5.13 / 73.6 ± 4.42	55.04 ± 1.54 / 65.72 ± 1.75
model-garden-lms/bert-base-token-dropping-finewebs-951k	66.87	88.81 ± 0.68 / 88.64 ± 0.54	57.44 ± 1.39 / 56.85 ± 2.09	50.91 ± 5.08 / 72.22 ± 4.2	54.63 ± 1.3 / 65.43 ± 1.43
model-garden-lms/bert-base-token-dropping-finewebs-901k	68.01	88.98 ± 0.64 / 88.67 ± 0.55	57.79 ± 1.31 / 58.91 ± 1.85	54.25 ± 6.3 / 75.73 ± 3.54	54.4 ± 0.72 / 65.31 ± 1.01
model-garden-lms/bert-base-token-dropping-finewebs-851k	67.97	88.9 ± 0.7 / 88.81 ± 0.54	58.0 ± 1.02 / 58.73 ± 1.8	54.04 ± 2.61 / 74.89 ± 2.07	54.75 ± 1.08 / 65.66 ± 1.26
model-garden-lms/bert-base-token-dropping-finewebs-801k	67.80	88.95 ± 0.7 / 88.73 ± 0.58	57.71 ± 1.43 / 60.5 ± 1.69	50.95 ± 6.3 / 74.16 ± 3.2	55.24 ± 1.37 / 66.13 ± 1.24
google-bert/bert-base-cased	62.26	87.39 ± 0.79 / 87.11 ± 0.66	54.49 ± 1.36 / 53.22 ± 1.15	52.08 ± 2.13 / 74.52 ± 1.31	38.63 ± 2.1 / 50.68 ± 1.87
google/electra-base-discriminator	69.26	87.82 ± 0.69 / 86.83 ± 0.62	62.3 ± 1.12 / 55.93 ± 0.67	62.61 ± 1.21 / 80.85 ± 0.59	52.51 ± 0.86 / 65.2 ± 0.85
FacebookAI/roberta-base	68.96	90.35 ± 0.23 / 90.14 ± 0.2	60.95 ± 1.4 / 57.52 ± 1.97	50.64 ± 1.69 / 74.55 ± 0.9	57.82 ± 1.35 / 69.68 ± 1.02

Our pretrained Token Dropping BERT model shows only a strong performance over the original BERT model. All detailed results can be found in this dataset repository.

❤️ Acknowledgements

This repository is the outcome of the last two years of working with TPUs from the awesome TRC program and the TensorFlow Model Garden library.

Made from Bavarian Oberland with ❤️ and 🥨.