Generate Mathematical Code Accompanied with Reasoning Steps

Step 1: get path to the files to be annotated

To get paths to the files to be annotated, run:

python data_processing/mathematical_code/get_files.py

This script creates file_names.json.

Step 2: start server

Get the Docker from text-generation-inference, and run deploy.sh in the Docker environment to start the server hosting Mixtral-8x7B-Instruct. Run:

bash data_processing/mathematical_code/deploy.sh

Step 3: start generating mathematical code

On the same machine as the server, run process.sh. Replace START_IDX and INTERVAL with the index and the number of process you are running. For example, if you started 4 processes on four different machines, then the INTERVAL is 4, and START_IDX is 0, 1, 2, 3 respectively.

bash data_processing/mathematical_code/process.sh START_IDX INTERVAL

Running this script results in annotation jsonl files saved under data/mathematical_code_orig.

Step 4: extract the conditions, expressions, results, and code snippets

Run:

python data_processing/mathematical_code/extract_to_blocks.py

This generates extracted files under data/mathematical_code_blocks.

Step 5: execute the code snippets

Run:

python data_processing/mathematical_code/execute_code.py

This results in files with execution results under data/mathematical_code_blocks_executed.

Step 6: filter based on the result of execution

Run:

python data_processing/mathematical_code/execution_result_filter.py

This generates files under data/mathematical_code_blocks_executed_filtered

Step 7: concatenate the reasoning step and code to form pretrain text

Run:

python data_processing/mathematical_code/convert_to_text.py

This generates the final files under data/mathematical_code