Features Required

[syncdoth]

• [x] 1. Need to write the prediction code: From the predicted logits, take argmax to get the corresponding class and get the corresponding sequence of tags. Then, mask the invalid tags.

• [x] 2. Evaluation code is written, but we need prediction code for the test data, and save it as a csv file containing strings of tags.

• [ ] 3. Training code is missing save and load feature.

• [x] 4. Implement different classification head options for BERT: LSTM, timedistributed dense, GRU, etc.

[syncdoth]

We need another evaluation code, which evaluates based on the ORIGINAL val target, not the preprocessed ones. Check for if the length matches: i.e.

1. no mistakes in masking,
2. max sequence length is enough.