Dataset, models, and implementation codes for the paper "MemAgent: A cache-inspired framework for augmenting conversational Web Agents with task-specific information"

The project aims to align user goals through multilevel conversations using the FastChat platform. By fine-tuning different language models, such as Vicuna 7B and Vicuna 13B, we facilitate effective communication and understanding between users and the system.

• Multilevel Conversations: The project leverages multilevel dialogues to ensure comprehensive goal alignment.
• FastChat Platform: We utilize the FastChat platform for efficient and context-aware interactions.
• Model Fine-Tuning: Fine-tuning of language models (e.g., Vicuna 7B, Vicuna 13B ...).
• Mind2Web Dataset: Our training data comes from the Mind2Web dataset, which provides diverse conversational examples for generalized tasks.

1. Clone this repository from here
2. Have the checkpoints from here.
3. Mind2Web dataset can be found in this repo "./mind2web dataset". From this dataset and with the help of gpt-4 (model: gpt-4-1106-preview), we generated our train data. Find the code in \gpt4datagen\DataGen directory
4. You may Fine tune with our given command script

scripts for fine tuning can be found in the ./scripts directory. The scripts for the execution phase are in the Execution directory.

1. Initialize the FastChat system. Install dependencies
2. Run with the following command:

    python3 -m fastchat.serve.cli --model-path path_to_directory  

3. You can also run inferencing with a specific checkpoint.

• Vicuna Models: We fine-tuned Vicuna 7B model.
• BLEU_Score: We have calculated bleu score with sacrebleu.
• BERT_score

We have developed the data generation for the training purpose We have introduced RAG pipeline