Update README.md

README.md

@@ -1,4 +1,11 @@
 ---
+license: cc-by-4.0
+pretty_name: HETUS Harmonized Time Use (Eurostat TSV + ESMS metadata)
+task_categories:
+- tabular-classification
+- text-classification
+language:
+- en
 dataset_info:
 - config_name: files
   features:
@@ -144,4 +151,127 @@ configs:
   - split: train
     path: observations/train-*
   default: true
+tags:
+- sociology
+- time-use
+- europe
+- EU
+size_categories:
+- 1M<n<10M
 ---
+
+# Dataset Card: HETUS Time Use (Converted from Eurostat TSV)
+
+## Dataset Summary
+
+This dataset contains normalized tabular observations converted from Eurostat HETUS TSV files, plus flattened SDMX ESMS metadata.
+
+The conversion is performed by `db-hf-normalization.py` and preserves raw TSV tokens to prevent information loss during normalization.
+
+## Source Data
+
+- Original source: Eurostat HETUS database exports (TSV + SDMX XML metadata)
+- Local source folder: `hetus/`
+- Waves currently included: `2000_2010`, `2020`
+
+## Conversion Method
+
+The converter:
+
+1. Reads TSV files as tab-separated raw text (no NA auto-coercion).
+2. Expands packed dimension keys from the first column.
+3. Melts year/time columns into long format.
+4. Preserves exact raw source tokens for auditability:
+   - `source_key_raw`
+   - `time_period_raw`
+   - `observation_cell_raw`
+5. Creates normalized fields for analysis:
+   - `time_period`, `time_period_year`
+   - `observation`, `observation_value`, `duration_minutes`, `status_flag`
+6. Validates strict lossless mapping per file (1:1 source cell preservation).
+
+## Splits
+
+### `observations`
+
+Long-format observations.
+
+Core provenance fields:
+
+- `dataset_id`
+- `wave`
+- `source_file`
+- `metadata_file`
+- `source_row_index`
+- `source_key_raw`
+- `time_period_raw`
+- `observation_cell_raw`
+
+Normalized fields:
+
+- `time_period`
+- `time_period_year`
+- `observation_raw`
+- `observation`
+- `observation_value`
+- `duration_minutes`
+- `status_flag`
+- `is_missing`
+
+Dimension fields (vary by source table, unioned into one schema):
+
+- Examples: `freq`, `unit`, `sex`, `age`, `geo`, `acl00`, `acl18`, ...
+
+### `files`
+
+Per-source-file conversion and validation report:
+
+- `rows_wide`
+- `rows_long`
+- `rows_expected_long`
+- `lossless_validated`
+- dimensions/time-period summaries
+
+### `metadata`
+
+Flattened SDMX ESMS metadata attributes with:
+
+- attribute id/path
+- HTML value (`value_html`)
+- plain-text value (`value_text`)
+
+## Data Loss / Integrity Statement
+
+For each TSV file, conversion enforces:
+
+- Exact row cardinality: `rows_long == rows_wide * number_of_time_period_columns`
+- Exact cell-level preservation of raw source tuples:
+  `(source_row_index, source_key_raw, time_period_raw, observation_cell_raw)`
+
+If any mismatch is detected, conversion fails for that file.
+
+## Known Caveats
+
+- The first TSV column stores packed dimensions separated by commas. The converter keeps both:
+  - raw packed key (`source_key_raw`)
+  - expanded normalized dimensions
+- Whitespace around values may be normalized in analytical columns, but the raw fields remain preserved.
+
+## Intended Uses
+
+- Comparative time-use analysis across countries, years, demographic strata, and activity codes.
+- Reproducible ETL pipelines where raw-source lineage is required.
+
+## Out-of-Scope Uses
+
+- Any individual-level inference (dataset is aggregate tabular statistics).
+
+## Citation
+
+Please cite Eurostat HETUS and include repository/version information for this converted dataset.
+
+## Dataset Version
+
+- Converter script: `db-hf-normalization.py`
+- Conversion date: {{YYYY-MM-DD}}
+- Repository commit: {{GIT_COMMIT_SHA}}