scancode-results-analyzer helps detect faulty scans in scancode-toolkit by using statistics and nlp modeling, among other tools, to make Scancode better.

How to Use

Refer quickstart section in scancode-results-analyzer documentation, for installing and getting started.

Link to Code

Then in NLPModelsPredict class, function predict_basic_lic_class uses this classifier to predict sentances as either valid license tags or false positives.

Limitations and Bias

As this model is a fine-tuned version of the bert-base-cased model, it has the same biases, but as the task it is fine-tuned to is a very specific task (license text/notice/tag/referance) without those intended biases, it's safe to assume those don't apply at all here.

Training and Fine-Tuning Data

The BERT model was pretrained on BookCorpus, a dataset consisting of 11,038 unpublished books and English Wikipedia (excluding lists, tables and headers).

Then this bert-base-cased model was fine-tuned on Scancode Rule texts, specifically trained in the context of sentence classification, where the four classes are

- License Text
- License Notice
- License Tag
- License Referance

Training Procedure

For fine-tuning procedure and training, refer scancode-results-analyzer code.

Link to Code

In NLPModelsTrain class, function prepare_input_data_false_positive prepares the training data.

In NLPModelsTrain class, function train_basic_false_positive_classifier fine-tunes this classifier.

Model - BertBaseCased (Weights 0.5 GB)
Sentence Length - 32
Labels - 4 (License Text/Notice/Tag/Referance)
After 4 Epochs of Fine-Tuning with learning rate 2e-5 (60 secs each on an RTX 2060)

Note: The classes aren't balanced.

Eval Results

Accuracy on the training data (90%) : 0.98 (+- 0.01)
Accuracy on the validation data (10%) : 0.84 (+- 0.01)

Further Work

Apllying Splitting/Aggregation Strategies
Data Augmentation according to Vaalidation Errors
Bigger/Better Suited Models

ayansinha/lic-class-scancode-bert-base-cased-L32-1

lic-class-scancode-bert-base-cased-L32-1

Intended Use