---
license: mit
task_categories:
- text-generation
- token-classification
language:
- en
tags:
- finance
- xbrl
- numeric-tagging
- concept-tagging
- fintagging
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test.parquet
---

# FinTagging Context-Aware Concept Tagging Test Set

This dataset combines the table-context and text-context test splits derived
from `FinTagging_800_200_HF`. It is intended for the second tagging step: given
an extracted numeric entity, datatype, local context, and original context,
predict the valid XBRL concept tag or tags.

Identical inputs are grouped. If the same input maps to multiple valid concepts,
the target is a JSON list containing all ground-truth concepts.

## Splits

| Split | Rows | Source occurrences | Table rows | Text rows | Multi-tag inputs |
|---|---:|---:|---:|---:|---:|
| test | 2,509 | 2,520 | 2,341 | 168 | 11 |

## Columns

- `source_sample_idx`: original source row index.
- `context_id`: original context identifier.
- `split`: always `test`.
- `input_type`: `table` or `text`.
- `input`: JSON string containing the model input fields.
- `input_fields`: structured version of `input`.
- `output`: JSON list of ground-truth XBRL concept tags.
- `ground_truth_concepts`: structured version of `output`.
- `ground_truth_count`: number of unique valid concepts for the input.
- `source_entity_indices`: original `numeric_entities` indices grouped into this row.
- `source_match_statuses`: first-stage parser match statuses for audit only.
- `source_occurrence_count`: number of source entity occurrences grouped into this row.

## Input Format

For table inputs, `input_fields` contains:

```json
{"numeric_entity": "...", "datatype": "...", "row_context": "...", "column_context": "...", "original_context": "..."}
```

For text inputs, `input_fields` contains:

```json
{"numeric_entity": "...", "datatype": "...", "sentence_context": "...", "original_context": "..."}
```