[WIP] Benchmark

[Alleria1809]

The previous PR is rather messy. In case it cause trouble for you to use the experiment script, I recreated my branch based on main and make a new PR.

[Alleria1809]

To test the model's performance, simply run the experiment and input configurations.

The uploaded data is randomly sampled from the paper's 7400+ dev dataset.
I tested the paper also on this dataset.

Note:
In the bottom performance section,
random 10 means the first 10 data from the uploaded hotpot_dev_v1_simplified_random_100.json.
first 10 records in the dataset means the first 10 records from the original 7400+ dev dataset.

I have the average steps because I get that from the react agent. But I didn't include this in this version of hotpotqa code.

TODO:

• Test on other models once the agent is updated
• Think about optimization

[liyin2015]

merge now, later will work on improvement

[liyin2015]

lets merge for now, will improve later