docs/schema.md

logo-lab/dlte per-config schema reference

Two configs:

• manifests: lightweight metadata as one parquet stream with a unified (record_type, record_json) schema.
• lake: full table content (parents + fragments + distractors) with typed columns + provenance keys.

manifests config

Single split (train), 109,336 records.

Column	Type	Notes
record_type	string	one of: parent, fragment, lake_table, split_assignment, query_task, ckan_distractor, fragmentation_config, gt_validation
record_json	string	JSON-encoded record content (variable schema per record_type)

record_type counts

record_type	rows	source
lake_table	47,772	manifests/lake_manifest.jsonl
ckan_distractor	36,740	manifests/ckan_distractor_ids.txt (one row per id)
fragment	16,548	manifests/fragments_manifest.jsonl
query_task	5,516	ground_truth/query_tasks.jsonl
parent	1,379	manifests/parents_filtered.jsonl
split_assignment	1,379	manifests/splits.json (one row per (parent_id, split))
fragmentation_config	1	config/fragmentation.json
gt_validation	1	ground_truth/gt_validation.json

Record schemas (after json.loads(record_json))

• parent: {parent_id, dataset, csv_path, csv_stem, table_id_in_embeddings, n_rows, n_cols, key_col, key_col_idx, key_uniqueness, key_non_null_ratio, key_is_text}
• fragment: {table_id, parent_id, split, role, fragment_type, noise_tier, replicate_id, csv_path, n_rows, n_cols, key_col_fragment, key_col_parent}
• lake_table: {table_id, source, parent_id, fragment_type, noise_tier, replicate_id, split, csv_path, n_rows, n_cols} — the full lake of 47,772 (= 11,032 targets + 36,740 distractors)
• split_assignment: {parent_id, split} (split ∈ {train, dev, test})
• query_task: {query_table_id, parent_id, split, noise_tier, replicate_id, relevant: [{table_id, relation}, ...]}
• ckan_distractor: {distractor_id}
• fragmentation_config: full config blob (tiers, noise specs)
• gt_validation: validation ground truth blob

lake config

Single split (train), 54,667 rows (= 1,379 parents + 16,548 fragments [5,516 query_seed + 5,516 target_union + 5,516 target_join] + 36,740 distractors). Sharded by byte budget (~1.5 GB raw per shard) so each parquet column page stays under the INT32 limit.

Column	Type	Notes
table_id	string	unique within the lake
kind	string	parent_tabfact / parent_wtq / target_union / target_join / query_seed / distractor
parent_id	string	empty for distractors
parent_source	string	tabfact / wtq / empty (distractor)
noise_tier	int32	0/1/2/3 for fragments; -1 for parents/distractors
fragment_type	string	seed / union / join / empty
split	string	train / dev / test / empty (distractors are unsplit)
csv_text	large_string	raw CSV content of the table
n_rows	int32	row count (excluding header)
n_cols	int32	column count

Counts by kind

kind	count	provenance
parent_tabfact	989	TabFact (Chen et al., ICLR 2020)
parent_wtq	390	WikiTableQuestions (Pasupat & Liang, ACL 2015)
target_union	5,516	derived: row-blocks removed from parents
target_join	5,516	derived: column-blocks removed from parents
query_seed	5,516	derived: parent minus removed rows + columns
distractor	36,740	LakeBench-Srinivas Zenodo CKAN Subset pool
Total	54,667	—

Of these, 47,772 = target_union + target_join + distractor is the retrieval lake the pipeline searches over. The other 6,895 (1,379 parents + 5,516 query seeds) are released alongside as pipeline I/O.

Cross-config keys

• manifests.lake_table.record_json["table_id"] ↔ lake.table_id for kind ∈ {target_union, target_join, distractor}
• manifests.parent.record_json["parent_id"] ↔ lake.parent_id (and matches lake.table_id for kind ∈ {parent_*})
• manifests.split_assignment.record_json["parent_id"] ↔ lake.parent_id for parent rows

Stage-2 calibrated thresholds for the (Stage-1 backbone, Stage-2 column model) pairs evaluated in the paper are not included in this release; they are deterministic given the dev split and reproducible from the operator spec in the paper appendix.