---
license: cc-by-4.0
tags:
- msmarco
- pretokenized
- glove
- cl100k
---

# MS MARCO triplet-50 pre-tokenized for `jsanzolac/drifting-glove-distilled-r300`

Pre-tokenized from `sentence-transformers/msmarco-msmarco-distilbert-base-v3` (subset `triplet-50`).
Each row: query, positive passage, and up to 50 mined hard negatives, all encoded with
`tiktoken.cl100k_base` and mapped to GloVe row indices via `jsanzolac/drifting-glove-distilled-r300/glove_vocab.txt`.

**Schema:**
- `query_ids: list[int32]` — GloVe row indices, length ≤ 256
- `positive_ids: list[int32]` — length ≤ 256, ≥ 8
- `neg_ids: list[list[int32]]` — variable length per row, each negative ≥ 8

**Filter rules:**
- Row dropped if query is empty (after OOV drop) or positive < 8 tokens.
- Negatives < 8 tokens are dropped from the row's negative list (row kept).

**Counts:** `rows_in = 278737`, `rows_kept = 278737` (100.00%).

See `preprocessing_stats.json` for length distributions.