README.md

# Zero-Shot Hierarchical Text Classification Dataset

This dataset is prepared from the consolidated climate change solution extraction results.

## Task

Zero-shot hierarchical multi-label text classification. Each example contains:

- `text`: classification input text built from solution details and document context.
- `label_paths`: one or more hierarchical labels, e.g. `Long-term -> Mitigation and clean energy -> Pollution control and clean energy promotion`.
- `solution_detail_items`: normalized `solution_details` items with their mapped label paths.
- `labels`: flattened labels for models that do not consume paths directly.
- `metadata`: city, country, text type, actor, climate event, source URL, and raw labels.

## Files

- `test.jsonl`: all evaluation examples for zero-shot classification.
- `all.jsonl`: same examples as `test.jsonl`, provided as a neutral full dataset file.
- `dataset.csv`: tabular version for quick inspection.
- `taxonomy.json`: normalized phase/domain/category hierarchy for both `solution` and `solution_details`.
- `candidate_labels.txt`: flat candidate label list.
- `stats.json`: dataset statistics and unmapped source labels.

## Statistics

- Source file: `consolidated_results_1022_events.csv`
- Total examples: `2083`
- Parent labels: `2`
- Domain labels: `8`
- Leaf labels: `50`

## Suggested Zero-Shot Usage

Use `taxonomy.json` as the candidate label space. For hierarchical prediction, first predict the phase (`Long-term` or `Short-term`), then the solution domain, then restrict final category candidates to that branch. For models that support multi-label classification directly, evaluate against `label_paths` or the flattened `labels` field.