# 005 — Parquet Conversion & HuggingFace Dataset Card

**Date**: 2026-03-15
**Scope**: Convert 5 key TSV files to Parquet (ZSTD), add YAML dataset card for `datasets` library integration

---

## 1. Objective

Enable programmatic access via `datasets.load_dataset("Nacryos/ancient-scripts-datasets", "config_name")` by:
1. Converting the 5 largest TSV files to Parquet format (columnar, compressed, typed)
2. Adding a YAML frontmatter dataset card to `README.md` with named configs pointing to the Parquet files
3. Preserving all original TSV files as legacy backups (no data removed)

### Why Parquet?
- The HuggingFace `datasets` library v4.0 removed support for custom loading scripts (they cause `RuntimeError`). YAML-based configs pointing to Parquet files are the recommended replacement.
- Parquet enables: columnar reads (load only needed columns), predicate pushdown (filter without full scan), proper null handling, typed schemas.
- TSV sentinel values (`-` for missing data) are converted to proper nulls.

## 2. Scripts Used

| Script | Lines | Purpose |
|--------|-------|---------|
| `scripts/convert_to_parquet.py` | ~100 | Read TSV via pyarrow, force all columns to string on initial read, convert `-` sentinels to null, cast typed columns (int/float), write Parquet with ZSTD compression |

### Key Implementation Details
- **Type safety**: All columns initially read as `pa.string()` to avoid pyarrow type inference issues (e.g., Score column auto-detected as `double` conflicting with `-` sentinel)
- **Null handling**: All `-` values converted to proper Parquet nulls via `pc.if_else(pc.equal(col, "-"), None, col)`
- **Compression**: ZSTD (Zstandard) — best compression ratio for this data profile
- **Column casting**: Int columns (Tree_Distance, MRCA_Depth) cast to `pa.int32()`, float columns (Score) cast to `pa.float64()`

## 3. Data Sources

No new external data. This change only converts existing TSV files already in the dataset.

### Input Files (preserved as legacy)
| File | Rows | TSV Size |
|------|------|----------|
| `cognate_pairs_inherited.tsv` | 22.9M | 2.2 GB |
| `cognate_pairs_similarity.tsv` | 465K | 49.9 MB |
| `cognate_pairs_borrowing.tsv` | 17K | 1.9 MB |
| `phylo_pairs.tsv` | 386K | 24.7 MB |
| `languages.tsv` | 1,177 | 39.3 KB |

## 4. Source Reputability

N/A — format conversion only, no new external data. All data integrity preserved from source TSVs.

## 5. Methodology

### Conversion Process
1. Read TSV with `pyarrow.csv.read_csv()`, forcing all columns to `pa.string()` schema
2. For each column, replace `-` sentinel with `null`
3. Cast numeric columns to proper types (`int32`, `float64`)
4. Write Parquet with `pq.write_table(table, path, compression='zstd')`

### YAML Dataset Card
Added to `README.md` frontmatter:
```yaml
configs:
  - config_name: cognate_pairs_inherited
    data_files:
      - split: train
        path: data/training/cognate_pairs/cognate_pairs_inherited.parquet
  - config_name: cognate_pairs_borrowing
    data_files:
      - split: train
        path: data/training/cognate_pairs/cognate_pairs_borrowing.parquet
  # ... (5 total configs)
```

### Why NOT a Loading Script
HuggingFace `datasets` v4.0 (released late 2025) removed custom loading script support entirely. Any repo with a `*.py` loading script triggers `RuntimeError`. YAML-based configuration is the official replacement.

## 6. Tests Performed

- **Row count verification**: Each Parquet file verified to have identical row count to source TSV
- **Schema verification**: Column names and types match expected schema
- **Random sample verification**: 10 random rows from each Parquet file compared byte-for-byte against corresponding TSV rows
- **Null handling**: Verified `-` sentinels are proper nulls (not empty strings) in Parquet
- **LFS integrity**: Verified no LFS pointer files were accidentally converted (detected and re-pulled `cognate_pairs_borrowing.tsv` which was a 3-line LFS pointer)
- **HuggingFace API verification**: Confirmed `https://huggingface.co/api/datasets/Nacryos/ancient-scripts-datasets` returns all 5 configs

## 7. Cross-Referencing

- Parquet file sizes verified to be strictly smaller than TSV (compression working correctly)
- Column count matches: 14 columns for cognate pairs, 9 for phylo_pairs, 5 for languages
- HuggingFace Dataset Viewer activated and showing data correctly

## 8. Output Summary

| File | TSV Size | Parquet Size | Compression Ratio |
|------|----------|-------------|-------------------|
| `cognate_pairs_inherited.parquet` | 2.2 GB | 23.6 MB | 93.8x |
| `cognate_pairs_similarity.parquet` | 49.9 MB | 5.0 MB | 10.0x |
| `cognate_pairs_borrowing.parquet` | 1.9 MB | 487 KB | 4.0x |
| `phylo_pairs.parquet` | 24.7 MB | 416 KB | 59.3x |
| `languages.parquet` | 39.3 KB | 13.7 KB | 2.9x |
| **Total** | **2.28 GB** | **29.5 MB** | **77.2x** |

## 9. Legacy Preservation

All original TSV files are preserved in their original locations. The Parquet files are stored alongside them:
- `data/training/cognate_pairs/cognate_pairs_inherited.tsv` (original, kept)
- `data/training/cognate_pairs/cognate_pairs_inherited.parquet` (new, added)

No data was removed or modified. The Parquet files are a read-optimized mirror.

## 10. SDK Package

A companion Python SDK (`ancient-scripts-data`) was also created in a separate repo (`Project-Phaistos/ancient-scripts-datasets-NEW`) to provide typed APIs for accessing this data. See that repo's README for API documentation.