Adding arc_easy, arc_challenge, mutual, mutual_plus

[yzpang]

After some discussion, we decided to implement different versions of a task (e.g., ARC) as two separate tasks. All the tasks in this PR (arc_easy, arc_challenge, mutual, mutual_plus) have the multiple choice formats.

To make things easier I used colabs to experiment. Here are the scripts: arc_easy, arc_challenge, mutual, mutual_plus.

Accuracies with RoBERTa-base:

• arc_easy: around 0.553
• arc_challenge: around 0.355 (discussed in meeting; this is low but expected given that the task is difficult, and we didn't use information retrieval models which the top-performing models on the leaderboard uses)
• mutual: around 0.709
• mutual_plus: around 0.643

Papers

• ARC (AI2 reasoning challenge)
• Mutual (a dataset for multi-turn dialogue reasoning)

ARC caveats

• The scripts are a bit complicated given that in the dataset, sometimes the examples are labeled by "A" to "D" and sometimes "1" to "4". Sometimes, there are five choices instead of four.

Please let me know if you need me to edit anything / if you need more info.

[codecov]

Codecov Report

Merging #1206 into master will increase coverage by 0.21%.
The diff coverage is 69.87%.

@@            Coverage Diff             @@
##           master    #1206      +/-   ##
==========================================
+ Coverage   56.91%   57.12%   +0.21%     
==========================================
  Files         133      137       +4     
  Lines        9696     9862     +166     
==========================================
+ Hits         5518     5634     +116     
- Misses       4178     4228      +50     

Impacted Files	Coverage Δ
...cripts/download_data/datasets/hf_datasets_tasks.py	100.00% <ø> (ø)
jiant/tasks/constants.py	100.00% <ø> (ø)
jiant/tasks/evaluate/core.py	37.15% <ø> (ø)
jiant/tasks/lib/arc_challenge.py	62.22% <62.22%> (ø)
jiant/tasks/lib/arc_easy.py	62.22% <62.22%> (ø)
jiant/tasks/lib/mutual.py	77.77% <77.77%> (ø)
jiant/tasks/lib/mutual_plus.py	77.77% <77.77%> (ø)
jiant/tasks/retrieval.py	100.00% <100.00%> (ø)
... and 2 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 0b3dff5...67f4fe9. Read the comment docs.

[zphang]

Merge from master and we should be good to go.