Repunt new model + quantization

Zip archive contains params.json with training params (we need this for configure our model's network), weights (*.pt) and logs.
Just extract this archive in models/ directory.
Expected directory structure after extraction:

models
└── repunct-model-new
    ├── logs
    │   └── repunct-model-new_logs.txt
    ├── params.json
    └── weights
        ├── weights_ep6_9912.pt
        └── _quantized_weights_ep5_9912.pt

Model evaluation metrics

Best validation Acc: 9912204619275377

Confusion Matrix:
	[22035741   102978     4773]
	[  107477  2112209     2878]
	[    7320     1558  1478818]
O:
	Precision: 0.9948
	Recall: 0.9951
	F1 score: 0.995
COMMA:
	Precision: 0.9528
	Recall: 0.9503
	F1 score: 0.9516
PERIOD:
	Precision: 0.9949
	Recall: 0.994
	F1 score: 0.9944
COMMA + PERIOD:
	Precision: 0.9697
	Recall: 0.9679
	F1 score: 0.9688

Repunt model

This model works only in repunct-stable branch

Zip archive contains params.json with training params (we need this for configure our model's network), last 2 best weights (*.pt) and logs.
Just extract this archive in models/ directory.
Expected directory structure after extraction:

models
└── repunct-model
    ├── logs
    │   └── repunct-model_logs.txt
    ├── params.json
    └── weights
        ├── weights_ep4_9910.pt
        └── weights_ep5_9911.pt

Model evaluation metrics

Best validation Acc: 0.9832204226585259

Confusion Matrix:
  [21489384   144837    40369]
  [  188716  1890897     8915]
  [   33862     5237  1359430]

O:
    Precision: 0.9897
    Recall: 0.9915
    F1 score: 0.9906

COMMA:
    Precision: 0.9265
    Recall: 0.9054
    F1 score: 0.9158

PERIOD:
    Precision: 0.965
    Recall: 0.972
    F1 score: 0.9685

COMMA + PERIOD:
    Precision: 0.9422
    Recall: 0.9321
    F1 score: 0.9371