k8-dmi3eva/spiderology

What is it?

• Dataset for evaluation text-to-sql models based on Spider dataset.
• It is divided into 32 groups (testsets) of samples. Each group checks some models ability.
• For each group there is control group (simplyset) - the same samples, but without considering features.
• This is very similar to A/B-testing technic.
• All the simplified samples (from control group) are new.
• Test samples contains:
  1. Examples extracted from Spider (not only from "dev", but also from "train" and "train_other") Be careful, if you use it for training!
  2. Examples extracted from Spider, but corrected from ambiguity and errors
  3. Manually created (new) examples

Structure

Folders and files

• [all] - the hole dataset in one place
  • all/testset_all.json - the samples with features from all testsets
  • all/simplyset_all.json - the samples without features from all testsets
  • all/spiderology.json - the samples with features from all testsets and simplysets
  • all/tables_all.json - the schemes of all tables required for requests
  • all/databases - the SQLite-databases for all requests in testsets and simplysets
• [testsets] - folder with testsets calledby their name
  • testsets/X/testset_X.json - the samples with feature X. Structure of samples is below
  • testsets/X/simplyset_X.json - the samples from testset without feature X. Structure of samples is the same
  • testsets/X/tables_X.json - the schemes of all tables for testset and simplyset
  • testsets/X/databases_X.json - the SQLite-databases for all requests in testsets and simplysets X
  • testsets/X/markup_X.json - the simplysets and testsets in one table (in the form convenient for comparing)
• [experiments/predictions] - folder with testsets calledby their name
  • /experiments/predictions/spiderology_with_predictions.json - all samples with predictions of two considering models

Structure of testsets and simplysets

{
  "id": Unique ID of sample
  "db_id": Name od database, corresponds to folder with DB in the databases_X
  "query": SQL-query
  "query_toks": Tokens of SQL-query
  "query_toks_no_value": Tokens of SQL-query, filter value replaced by token 'value'
  "question": Natural language question  
  "question_toks": Tokens of natural language question 
  "sql": Structured SQL-query (in Spider format
  "tags": List of tags (features an sub-characteristics)
  "source": Variant of sources (see below)
  "parents_id": None for samples in testset or ID of simplified sample for simplification
  "is_simplification": 'true' for simplifications, 'false' for other samples
  "simplifications_tags": None for samples from testset, list of named simplified features for simplyset,
  "predictions": Dictionary with SQL-queries predicted by models (exists only in the folder "experiments"),
  "evaluation": Results of exact_matching, execution and human evaluation (exists only in the folder "experiments")
}

Sources

• SIMPLIFIED = Simplification of samples
• SPIDER_DEV = Samples from Spider test set ('dev')
• SPIDER_TRAIN = Samples from Spider train set ('train')
• SPIDER_TRAIN_OTHERS = Samples from Spider train set ('train_other')
• ADDED = Created samples with features (not from Spider)

Test-set description

Here: https://github.com/rology/spiderology?tab=readme-ov-file#test-set-description