Dialogue Classifier

This is a hierarchical classifier that can be trained to process the data in the dialogue format.

How to run

1. The data must be in .jsonl format, where each line is one dialogue. Each line has the folowing structure:

   {
       "id": "000",
       "turns": ["Dialogue turn #0", "Dialogue turn #1", "Dialogue turn #2"], 
       "labels": [0, 0, 1, 2, 0, 0, 2, 3]
   }

   Train, validation and test sets must be called train.jsonl, validation.jsonl and test.jsonl and place into data/json/{DATASET_NAME} directory, where {DATASET_NAME} is the name of the dataset.

2. Example command to run the training:

   python run_train.py --dataset_name MY_DATASET --model_name_or_path intfloat/e5-base-v2 \
   --output_dir saved_models/MY_DATASET-e5v2-fixed_steps-smoothl1-optimized --do_train --per_device_train_batch_size 4 --per_device_eval_batch_size 4 \
   --gradient_accumulation_steps 8 --evaluation_strategy steps --eval_steps 200 --max_steps 2000 --warmup_ratio 0.1 \
   --logging_strategy steps --logging_steps 25 --save_strategy steps --save_steps 200 --seed 0 --data_seed 0 \
   --word_encoder_layers_to_train ".*attention.*$" \
   --phrase_hidden_size 768 --phrase_intermediate_size 1536 --loss_reduction mean \
   --optim adamw_torch --learning_rate 0.00002

3. Example command to run the prediction on the test set:

   python predict_optimized.py --model_name saved_models/MY_DATASET-e5v2-fixed_steps-smoothl1-optimized/checkpoint-1000 --dataset_name MY_DATASET --output predictions/fixed/MY_DATASET-e5v2-fixed_steps-smoothl1-optimized-1000-test.json --device cuda --data_split test