Huggingface Albert [unofficial Version]

hugginface albert model and its tokenizer
The albert repo transfer weights from tf-hub module and can run examples/run_squad_albert.py to reproduce author's performance.
support for albert_base, albert_large, albert_xlarge

First, prepare your data

{Training data path} and {Testing data path}.

Current status

you can run the squad task for albert

example

python run_squad_albert.py --train_file {Training data path} \
--predict_file {Testing data path}  \
--model_type albert \ 
--output_dir ./example_do_lowercase_lamb \
--do_train --do_eval \ 
--overwrite_output_dir --do_lower_case \ 
--evaluate_during_training --warmup_steps 384  \ 
--save_steps 1000 --logging_steps 1000 --num_train_epochs 3 \

if you want albert_large or albert_xlarge, you need to download the pytorch_state_dict on here https://drive.google.com/drive/folders/1OZuoBw0cNkPLFXdoz4xTZrTXdRHO7hUI?usp=sharing
put on examples/pytorch_model_state_dict
set new a argument below as --model_dict_pretrain ./pytorch_model_state_dict/{model_file name}.
adjust the config file set to --config_pretrain ./config/{alber_xxx.json} then you can use it for big model.

The Model Performance

Version 1

Run on Squad 1.1 task (albert base)
EM 78.89 , F1 score 87.38

The paper report the performance is below:
EM 82.3 F1 score 89.3

xlarge paper report the performance is below Squad 1.1
EM 86.1 F1 score 92.6

My performance for xlarge
EM 84.52 F1 Score 91.61

Version 2

My Albert_xl2 on squad 1.1 the performance
EM 86.32 F1 Score 92.99 still training.......for 3 epoch !!!!!!

The paper report the version 2 for albert xlarge on Squad 1.1 is
EM 86.1 F1 Score 92.5

Performance On version 2 albert base (Squad 1.1)

Exact Match:

F1 Score: