Add some documentation

This commit is contained in:
hlillemark 2025-03-22 16:09:59 -07:00
parent 1cefe44a15
commit d6094eecd6

View file

@ -15,6 +15,16 @@ import glob
This script is intended to run the evaluation script multiple times and then automatically aggregate the
successful logs into a subfolder for each run, based on the success marked in the experiment folder. Then
at the end it will aggregate everything into a json file, ready for training.
Example usage:
python3 ./multi_data_collection_script.py --api vllm --model meta-llama/Meta-Llama-3-8B-Instruct --num_agents 2 --num_parallel 2 \
--tasks "tasks/crafting_tasks/test_tasks/tasks_2_agents.json:3" "tasks/crafting_tasks/test_tasks/tasks_3_agents.json:3"
Meaning run those two tasks each 2 times, with num agents. The results will be in
./full_run_logs_{date}
and ./successful_run_logs_{date}
Use the successful run logs to train the next model.
"""