retreival-generation-system

Code files

• main.py : main function for odqa/cqa chatbot
• module.py : chatbot module classes
• finetune.py : finetune function for t5/opt
• model_utils.py : helper functions for module
• data_utils.py : helper function/classes for training data
• reranker.py: ranker for response quality, loaded from main_fn
• entity_tracker.py: entity tracker for dialog(in progress)

Scripts

• push_all.sh : push codes on github
• run_chatbot.sh : script for running chatbot
  • For single turn QA, use --task 'odqa'. For Conversational QA, use --task 'cqa' in script.
  • To serve OPT as generator, use --gen_model_path "facebook/opt-1.3b" and --model_name "opt" ; to use Flan-T5, change them to --gen_model_path "google/flan-t5-large" and --model_name t5
  • To load fine-tuned weight file, add --gen_cp_path <path_to_weight>. Without this line the pretrained weight of the model will be loaded.
• run_train_opt.sh : script for finetune opt
  • Feel free to change:
    • --device: The GPU# for training.
    • --logger_path: Path to save training log.
    • --saved_model_path: Path to save fine-tuned weight.
    • --outer_batch_size: "Batch size" for Gradient
    • --dataset_path: Path to converted training dataset, same path as --saved_dataset_path from run_convert_dataset.sh.Default ./dataset/train_data
    • --epochs: fine-tuning epochs
    • --lr: learning rate
  • Since training data is very large, the Gradient accumulation is applied in the code. Stay --batch_size 1 to avoid OUT_OF_MEMORY issue. Change --outer_batch_size to adjust the "batch size" for training.
• run_train_t5.sh : script for finetune t5
  • Same as run_train_opt.sh
• run_convert_dataset.sh: script to convert dataset generated from GPT-3 to the format used for training
  • --original_json_path: Path to GPT-3 generated data. The default dataset is GPT-3_semantic_search.json.
  • --converted_json_path: Path to converted Json format of dataset.
  • --saved_dataset_path: Path to save converted dataset

Dataset

• ece_rag_dataset_new : ECE120 course note dataset for DPR.
• train_data : ECE120 finetune dataset, converted from GPT-3_semantic_search.json
• GPT-3_semantic_search.json: GPT-3 generated dataset for training.

Issues

1. The context in training data is too large,truncate text may cause problem. (change the number in data_utils.py, line 214/247 to change the truncated size)
2. Max input Length for T5 is only 512, fine-tuning may work, but it may lead to accurate issues.
3. Entity Extractor for dialog in progress.