jeremycochoy/contrastive-training-small-bundles

# small_mixed_v1

Pre-shuffled, pre-mixed training bundle for the Small tier (42.5M params)
of the contrastive forecasting model family.

- **Total rows:** 27282687
- **Number of shards:** 2736

## Per-source row counts (source_id -> rows)

- 0 gift: 20250495 (74.2%)
- 1 wiki_hourly: 3715121 (13.6%)
- 2 wiki_daily: 1990244 (7.3%)
- 3 wiki_stl_residual: 530731 (1.9%)
- 4 wiki_stl_seasonal: 371512 (1.4%)
- 5 wiki_stl_trend: 159219 (0.6%)
- 6 synthetic: 265365 (1.0%)

## Schema

| Column | Type | Notes |
|---|---|---|
| | | Fixed-length window |
| | | 0=gift, 1..5=wiki sub-sources, 6=synthetic |
| | | Source-specific metadata |

## Shuffling and sampling

Globally shuffled via two-pass bucket shuffle (stage 3). Every output
shard is a statistically uniform random sample of the entire input.
Max mix-ratio delta between first and last shard: 0.009.

GIFT per-file sampling is byte-weighted. See the generating repo
() for details.

NaN rate: 0.06% all-NaN, 0.22% partial-NaN. Consumers should
forward-fill partial NaN and skip all-NaN rows.

Generated by (PRs #194-#202).