gift-pretrain-small

A small, single-source training bundle plus a mirror of the GIFT-Eval evaluation set, packaged together for convenience. Built by uniformly sampling 10 series from every sub-dataset of Salesforce/GiftEvalPretrain, then cropping each selected series into non-overlapping windows of length 1025 and globally shuffling the result.

This bundle exists as a sanity / smoke-test corpus before scaling up to the full GIFT-Eval Pretrain. Every sub-dataset of GIFT-Eval Pretrain (~152 directories at the time of build) contributes 10 series; nothing is dropped to byte-weight underflow.

Layout

.
├── small_v1/
│   ├── shard_NNNNN.parquet   ← gift-only training shards
│   └── manifest.json         ← total rows, num shards, source_counts
├── eval/                     ← Salesforce/GiftEval mirror (bytes-as-is)
│   ├── m4_daily/
│   ├── ett1/
│   └── ...
└── README.md

The eval/ tree mirrors Salesforce/GiftEval exactly so you can point the official gift-eval tooling at it without translation.

Schema (small_v1/shard_*.parquet)

Column	Type	Notes
series	list<float32>[1025]	Fixed-length non-overlapping window
source_id	uint8	Always 0 (gift) — bundle is single-source
meta	string	Original item_id from the source arrow file

Compression: zstd, row group size 10_000.

Sampling

For each sub-dataset (top-level directory) in Salesforce/GiftEvalPretrain, pick the smallest arrow file in that directory and sample K=10 rows from it uniformly without replacement. From each selected row, emit every non-overlapping window of length 1025. The output is globally shuffled at row granularity via the two-pass bucket shuffle used by the main contrastive-training bundles.

Sampling-bias notes:

• One arrow file per sub-dataset. Multi-shard sub-datasets like buildings_900k/ split a single logical dataset across many arrow files of similar size; sampling K series within one of them is a valid (slightly biased) sample of the whole. Picking the smallest file minimizes download bandwidth.
• No byte-weight underflow. The byte-weighted multinomial used by the full contrastive-training pipeline can drop small sub-datasets at low budget; this bundle deliberately gives every sub-dataset an equal seat at the table.
• Coverage of arrow files within a sub-dataset is intentionally low. This is a "small" bundle, not an unbiased sample of the whole 975 GB corpus.

Reproducing

The pipeline lives in scripts/training_data_prep/ of the generating repo. Build script:

python -m training_data_prep.build_gift_only_bundle \
    --output-dir /path/to/out \
    --series-per-subdataset 10 \
    --max-workers 4

Upload (two calls of the existing stage4_upload_hf.py, one for the bundle and one recursive for eval/):

python -m training_data_prep.stage4_upload_hf \
    --repo-id <user>/gift-pretrain-small \
    --bundle-dir /path/to/out/final \
    --path-in-repo small_v1 \
    --create-repo

python -m training_data_prep.stage4_upload_hf \
    --repo-id <user>/gift-pretrain-small \
    --bundle-dir /path/to/gift-eval-data \
    --path-in-repo eval \
    --recursive