SQuAD update

[zphang]

1. Fixes SQuAD v2 task config to properly supply kwargs
2. Update SQuAD tokenization logic based on Fix tokenization in SQuAD for RoBERTa, Longformer, BART huggingface/transformers#7387 (and subsequently refactored in [tests|tokenizers] Refactoring pipelines test backbone - Small tokenizers improvements - General tests speedups huggingface/transformers#7970)

The summary for 2 is that previously (including in the Hugging Face implementation), the context was being tokenized per-word rather than as a string. This is problematic for tokenizers that treat start-of-word tokens differently.

Although the cited PR states that there is a large performance improvement from this fix, I did not observe this myself in my testing (both in jiant as well as in the transformers examples). Both versions appeared to perform comparably.

For jiant users, this will impact tokenization, but training a fresh model for either old or new tokenization should both work. I recommended deleting and recreating any SQuAD-based caches. However, if you consistently use only the old tokenization caches, that should be fine too.

[codecov]

Codecov Report

Merging #1232 (99fc599) into master (e2e85c9) will decrease coverage by 0.01%.
The diff coverage is 20.00%.

@@            Coverage Diff             @@
##           master    #1232      +/-   ##
==========================================
- Coverage   56.57%   56.56%   -0.02%     
==========================================
  Files         147      147              
  Lines       10578    10582       +4     
==========================================
+ Hits         5985     5986       +1     
- Misses       4593     4596       +3     

Impacted Files	Coverage Δ
...t/scripts/download_data/dl_datasets/files_tasks.py	7.49% <ø> (ø)
jiant/tasks/lib/templates/squad_style/core.py	30.71% <20.00%> (-0.09%)	⬇️

---

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 e2e85c9...99fc599. Read the comment docs.

[jeswan]

Thoughts on moving tokenizer specific wrapping constants to a dedicated file (ext/tokenizers/constants.py or similar)? Worried about the logic being scattered throughout files.

[zphang]

This constant comes from Transformers: https://github.com/huggingface/transformers/blob/969859d5f67c7106de4d1098c4891c9b03694bbe/src/transformers/data/processors/squad.py#L17

Unfortunately, it comes in a later version than what we're currently requiring, so we can't currently import it (unless we bump the version, I'd like to wait until after Fall to do it because of the v3.0.2 issue). The upside is that it's from the same file that all the rest of the SQuAD code comes from, so it'll be consistent for anyone looking up update the SQuAD implementation in the future.