feat(datasets): simplify load flow from hf datasets with no rb format

[frascuchon]

With this PR you can convert hf datasets to rubrix datasets as follow:

import datasets
amazon_reviews = datasets.load_dataset("amazon_reviews_multi", "es", split="train")
rb_ds = rb.DatasetForTextClassification.from_datasets(
    amazon_reviews, 
    inputs=["review_title","review_body"], 
    annotation="stars"
)
# Here you can just log the dataset
rb.log(rb_ds, name="amazon-reviews")

We can define similar approach for metadata or text for the other tasks, or tags for token classification

@dcfidalgo @dvsrepo any suggestion?

TODO:

• Include tests for text2text
• Fix failing tests once hf bug is fixed:

import datasets
ds = datasets.Dataset.from_dict({ "a": [None, [1,2], None, [3,4]]})
ds.map(lambda x: { "b": "b"}).to_pandas() # None values in a becomes empty lists

See this PR huggingface/datasets#3676

[codecov]

Codecov Report

Merging #1234 (52421b5) into master (0935db7) will decrease coverage by 0.38%.
The diff coverage is 76.31%.

@@            Coverage Diff             @@
##           master    #1234      +/-   ##
==========================================
- Coverage   94.68%   94.29%   -0.39%     
==========================================
  Files         127      127              
  Lines        5470     5573     +103     
==========================================
+ Hits         5179     5255      +76     
- Misses        291      318      +27     

Flag	Coverage Δ
pytest	94.29% <76.31%> (-0.39%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
src/rubrix/client/models.py	88.88% <48.38%> (-8.01%)	⬇️
src/rubrix/client/datasets.py	95.46% <86.74%> (-2.74%)	⬇️

---

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 7bf10d4...52421b5. Read the comment docs.

[dvsrepo]

With this PR you can convert hf datasets to rubrix datasets as follow:

import datasets
amazon_reviews = datasets.load_dataset("amazon_reviews_multi", "es", split="train")
rb_ds = rb.DatasetForTextClassification.from_datasets(
    amazon_reviews, 
    inputs=["review_title","review_body"], 
    annotation="stars"
)
# Here you can just log the dataset
rb.log(rb_ds, name="amazon-reviews")

We can define similar approach for metadata or text for the other tasks, or tags for token classification

@dcfidalgo @dvsrepo any suggestion?

This is looking good!

For text it should be straightforward, for metadata there are some dataset which contain json objects in some fields.

For token classification we need at least tokens and tags and then generate stand-off annotations with that, right?

[dcfidalgo]

Nice shortcut, will definitely be helpful. Just as a side note, for token classification datasets on the Hub, often only tokens are provided, so we must reconstruct the text input for the TokenClassificationRecords.

[frascuchon]

Nice shortcut, will definitely be helpful. Just as a side note, for token classification datasets on the Hub, often only tokens are provided, so we must reconstruct the text input for the TokenClassificationRecords.

Yes, we can recompose text field in terms of tokens field by assuming white spaces between tokens

[frascuchon]

For token classification we need at least tokens and tags and then generate stand-off annotations with that, right

Agree with this. Although it would also take that logic to the creation of a token classification record

record = TokenClassificationRecord(tokens=....,tags=.....)

[frascuchon]

I've included a workaround to fix the test. The token classification model will try to build the text field from tokens if no char-level span is provided.

But in general, operating with None annotation/prediction lists could raise errors related to issue: huggingface/datasets#3676

Anyway, we can go ahead with this feat. even if issue is not resolved since most of hf datasets will provide annotations

[dcfidalgo]

💯 Very nice!