README.md

# Evidence-to-Taxonomy Retrieval: FinTagging / US-GAAP 2024 splits

The two splits used in *Factorized Hypothesis Search for Evidence-to-Taxonomy Retrieval*.
Each record is one **fact**: a value observed in a filing, together with the source context it
came from, and the US-GAAP concept(s) a human tagger assigned to it. The task is to retrieve
that concept from the 17,388-concept taxonomy given only the context and the value — there is
no natural-language query in the input.

| file | facts | source contexts | table / text facts | distinct gold concepts | md5 |
|---|---|---|---|---|---|
| `data/test.jsonl` | 2,509 | 191 | 2,341 / 168 | 388 | `70b4a0f4dc64ada13c47bdb657a26ce9` |
| `data/dev.jsonl` | 661 | 70 | 566 / 95 | 180 | `8bfd4c17373d2727187aa10d0c756b8c` |

**How each split was used.** Every number reported in the paper is computed on `test.jsonl`.
`dev.jsonl` is the configuration-selection sample: the number of hypotheses $J$ and the score
weight $\beta$ were chosen on it, and it is also the sample behind the appendix tables marked
`[DEV]`. No test fact influenced any configuration choice.

**Splits are disjoint at the context level.** All facts from one source context stay together,
so a context never appears in two splits; the two files share **zero** `context_id` values.
`dev.jsonl` carries `split: "train"` because the development sample is drawn from the training
split of the underlying benchmark, not from a separate pool — the field records provenance, not
the role the file plays here.

## Record schema

Both files use the same schema, except that `dev.jsonl` adds `fact_id` (its index within the
sample). Fields:

| field | type | meaning |
|---|---|---|
| `context_id` | int | source context; the unit to cluster on when bootstrapping |
| `source_sample_idx` | int | index of the context in the underlying benchmark |
| `split` | str | provenance split of the source context |
| `input_type` | str | `table` or `text` |
| `input` | str | JSON string of `input_fields`, kept for loaders that want one flat column |
| `input_fields` | dict | the parsed context: `numeric_entity`, `datatype`, and the row/column or sentence context |
| `ground_truth_concepts` | list[str] | gold US-GAAP concept tags |
| `ground_truth_count` | int | length of the above |
| `output` | str | JSON string of `ground_truth_concepts` |
| `source_entity_indices` | list[int] | where the value occurs in the source context |
| `source_match_statuses` | list[str] | how the value was located: `exact_cell`, `exact_sentence`, `ambiguous_cell`, `ambiguous_sentence`, `unmatched` |
| `source_occurrence_count` | int | number of occurrences found |
| `fact_id` | int | **dev only**: index within the development sample |

`input` / `output` duplicate `input_fields` / `ground_truth_concepts` as JSON strings. They are
kept because the runs read them; a loader should prefer the structured fields.

## The retrieval index

`taxonomy/us_gaap_2024_enriched_retrieval.jsonl` (17,388 concepts, ~12 MB,
md5 `a33ecd6538a11b77d2d6d2fa18fff237`) is the index every reported retrieval number is measured
against — the `taxonomy_jsonl` path recorded in all 69 run manifests. One concept per line:

| field | type | meaning |
|---|---|---|
| `tag` | str | US-GAAP concept name, without the `us-gaap:` prefix |
| `type` | str | XBRL item type (`monetaryItemType`, `stringItemType`, …). This is the datatype the retriever filters on |
| `standard_label` | str | readable standard label |
| `documentation` | str | documentation label from `us-gaap-doc-2024.xml`; **present for 14,918 of 17,388 concepts** |
| `references` | list[str] | paragraph-level ASC references from `us-gaap-ref-2024.xml`; present for 11,888 |
| `retrieval_text` | str | `tag`, `standard_label` and `documentation` joined — **this string, and only this string, is what BM25 indexes** |

Two things a reader should not have to discover by experiment: `references` is carried but is
**not** part of `retrieval_text`, so it never enters the lexical score; and retrieval is
restricted to concepts declaring the fact's own `type`, falling back to the full index when the
taxonomy declares none. `taxonomy/build_summary.json` records the four US-GAAP 2024 source files
this was built from and the per-field coverage counts.

Concept counts by type (top): monetary 7,650, string 4,628, domain 1,915, textBlock 982,
enumerationSet 569, percent 463.

## Provenance

All three files are byte-identical copies (verified by md5) of the files the reported runs read:

- `data/test.jsonl` — `FinTagging_800_200_grounding_test_JSON/data/test.jsonl`, the path recorded
  in the `test_jsonl` field of all 53 run manifests behind the paper's tables.
- `data/dev.jsonl` — `runs_ags_coverage_pilot/qwen3_32b/sample_facts.jsonl`, the default
  `SAMPLE_PATH` of the configuration-selection scripts.
- `taxonomy/us_gaap_2024_enriched_retrieval.jsonl` — `retrieval_data/us_gaap_2024_enriched/`,
  the `taxonomy_jsonl` path recorded in all 69 run manifests.