# soc91-labels

Mirror of the R2 prefix `soc91-labels/` from `soc127-dedup` (Cloudflare R2) into a private
HF Dataset. Generated by `scripts/handoff/mirror_r2_sidecar_prefix.py` on
2026-05-23.

Each row in the source R2 parquets is preserved, with one additional column appended:
`source_shard_path` — the R2 object key the row came from. This lets you reconstruct
the per-shard view if needed.

## Counts

| | Source (R2) | Mirror (this dataset) |
|---|---|---|
| Files | 58,465 | 2,719 (concat into ~60 MB chunks) |
| Size | ~150 GB (approx, mixed schemas) | 169.94 GB |

## Loading

```python
from datasets import load_dataset
ds = load_dataset("HCAI-Lab/soc91-labels")
# or for a single chunk:
import pyarrow.parquet as pq
table = pq.read_table("data_00000.parquet")
```

## Schema

Schema matches the per-shard WebOrganizer sidecar parquet schema as defined in
`src/dolma/sidecar_manifest_fields.py:43-124` (data-attribution repo), plus the appended
`source_shard_path` column. The underlying classifier produced both an "old" schema
(label ids only) and a "new" schema (label ids plus per-label probability vectors);
the concatenated table promotes nullable columns so rows from both schemas coexist.

Per-row fields (subset; see source for the full list):

- `doc_id` — Dolma document UUID
- `weborganizer_topic_id`, `weborganizer_topic` — top-1 topic class (24 categories)
- `weborganizer_format_id`, `weborganizer_format` — top-1 format class (24 categories)
- `weborganizer_topic_confidence`, `weborganizer_format_confidence` — top-1 probabilities
- `weborganizer_topic_probabilities`, `weborganizer_format_probabilities` — full probability vectors (new-schema rows only; null on old-schema rows)
- `source_shard_path` — the R2 object key the row originated from

## Lineage

- SOC-91 (initial label generation)
- SOC-142 (inverted-quality-label fix at commit `3342baf`; affects the quality sidecars in
  `HCAI-Lab/soc139-quality-sidecars`, not the topic/format labels here)

## Notes

- The chunked-write/stream-upload pipeline succeeded and committed all 2,719 chunks.
- The original `write_readme()` step crashed (after upload) because the staging directory
  files were already deleted by the stream-upload pattern. This README was written by
  hand after the fact based on the script's final log output (`wrote 2719 chunks total`).