danitamayo's picture
Update README.md
8c40f69
|
raw
history blame
2.96 kB

Results after fine-tuning distilbert in 80% of 15189 instances (Model still under development)

20it [00:11, 1.85it/s]Train: wpb=2121, num_updates=20, accuracy=44.1, loss=0.00
50it [00:28, 1.76it/s]Train: wpb=2121, num_updates=50, accuracy=55.4, loss=0.00
100it [00:55, 1.88it/s]Train: wpb=2117, num_updates=100, accuracy=64.5, loss=0.00
200it [01:48, 1.85it/s]Train: wpb=2132, num_updates=200, accuracy=71.6, loss=0.00
300it [02:42, 1.88it/s]Train: wpb=2147, num_updates=300, accuracy=75.1, loss=0.00
380it [03:24, 1.86it/s]
Train: wpb=2142, num_updates=380, accuracy=76.9, loss=0.00
| epoch 000 | train accuracy=76.9%, train loss=0.00
| epoch 000 | valid accuracy=85.7%, valid loss=0.00\

20it [00:10, 1.85it/s]Train: wpb=2121, num_updates=20, accuracy=84.6, loss=0.00
50it [00:27, 1.77it/s]Train: wpb=2121, num_updates=50, accuracy=84.6, loss=0.00
100it [00:54, 1.87it/s]Train: wpb=2117, num_updates=100, accuracy=85.1, loss=0.00
200it [01:47, 1.86it/s]Train: wpb=2132, num_updates=200, accuracy=85.4, loss=0.00
300it [02:41, 1.88it/s]Train: wpb=2147, num_updates=300, accuracy=85.6, loss=0.00
380it [03:24, 1.86it/s]
Train: wpb=2142, num_updates=380, accuracy=85.8, loss=0.00
| epoch 001 | train accuracy=85.8%, train loss=0.00
| epoch 001 | valid accuracy=88.3%, valid loss=0.00

20it [00:10, 1.86it/s]Train: wpb=2121, num_updates=20, accuracy=87.3, loss=0.00
50it [00:27, 1.77it/s]Train: wpb=2121, num_updates=50, accuracy=87.0, loss=0.00
100it [00:54, 1.88it/s]Train: wpb=2117, num_updates=100, accuracy=87.2, loss=0.00
200it [01:47, 1.85it/s]Train: wpb=2132, num_updates=200, accuracy=87.2, loss=0.00
300it [02:41, 1.88it/s]Train: wpb=2147, num_updates=300, accuracy=87.2, loss=0.00
380it [03:23, 1.86it/s]
Train: wpb=2142, num_updates=380, accuracy=87.3, loss=0.00
| epoch 002 | train accuracy=87.3%, train loss=0.00
| epoch 002 | valid accuracy=89.3%, valid loss=0.00

We have to change the loss function... It seems to be a problem...

You can evaluate the performance of our model by writing the following example: "google chrome before 18. 0. 1025. 142 does not properly validate the renderer's navigation requests, which has unspecified impact and remote attack vectors."

The result, for each token, should be similar : ['B-vendor', 'B-application', 'B-version', 'I-version', 'I-version', 'I-version', 'I-version', 'I-version', 'I-version', 'I-version', 'I-version', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'B-relevant_term', 'O', 'O', 'O', 'O', 'O', 'O', 'B-relevant_term', 'B-relevant_term', 'O', 'O']

Different possible classes that are detected: ['I-update', 'I-version', 'B-programming language', 'B-relevant_term', 'B-parameter', 'I-relevant_term', 'B-vendor', 'B-function', 'B-version', 'B-hardware', 'I-application', 'B-os', 'O', 'B-cve id', 'B-update', 'I-edition', 'I-hardware', 'I-os', 'B-edition', 'B-application', 'B-language', 'B-file', 'B-method', 'I-vendor']