README

This project involves generating, processing, training, and testing a dataset using several scripts. The following steps outline the entire process:

If you don't want to generate your own dataset, please skip to step 3.

1. Generate Dataset using dataset_preparing

First, generate the dataset using the scripts available in the dataset_preparing directory. Each script is designed to generate a dataset with a specific number of values.

2. Filter Data using find_error_data.py

Next, filter the generated dataset using the find_error_data.py script. This script will process the JSON files to identify and filter out entries influenced by specific value dimensions.

3. Preprocess Data using preprocess.py

Once the data is filtered, preprocess the data using the preprocess.py script. This step prepares the data for training by performing necessary transformations and formatting. Set different encoder model using parser argument -e. Set different dataset using parser argument -m (including 'train' and 'test').

4. Train Models using CL_labeling_v2.py and ACCL_CJRN.py

With the preprocessed data, train the models using the CL_labeling_v2.py and ACCL_CJRN.py scripts. These scripts implement the training process for the Sequential Evaluation Network (SEN) with the Adaptive Complexity Curriculum Learning (ACCL) framework. Set different encoder model using parser argument -e.

5. Test Value-Module using inference_combine_results.py

After training the models, test the value-module using the inference_combine_results.py script. This script evaluates the model's performance and combines the results from different networks. Set different encoder model using parser argument -e.

6. Test Action Selection in Scenarios using action_choosing.py

Then, test the action selection in different scenarios using the action_choosing.py script. This script evaluates and ranks actions based on user preferences and the results from the trained models. Set different encoder model using parser argument -e.

7. Compare final results using human_study/compare.py

Finally, test model performance in action choosing with real human and other benchmark LLM models using human_study/compare.py. This script list out the similarity between real human data and models, including our ValuePilot and other large language models. Set different encoder model using parser argument -e.