README.md

---
pretty_name: GLINT training data
task_categories:
- text-retrieval
- text-ranking
language:
- en
tags:
- hard-negatives
- late-interaction
- multi-vector
- colbert
- retrieval
- maxsim
size_categories:
- 1M<n<10M
license: apache-2.0
---

# GLINT training data

The central artifact is a
seven-source hard-negative mixture mined with a **multi-vector late-interaction model**, rather
than a dense bi-encoder or BM25.

Candidate documents are retrieved under **MaxSim**: each query token is matched to its best
document token, and those best similarities are summed. This produces negatives that are hard
under the same token-aware scoring geometry GLINT uses at retrieval time.

- **Multi-vector mining model:**
  [lightonai/LateOn-unsupervised](https://huggingface.co/lightonai/LateOn-unsupervised)
- **Mining backend:** [PyLate](https://github.com/lightonai/pylate) +
  [WARP](https://github.com/jlscheerer/xtr-warp)
- **KD mixture:** 1,046,009 query rows × 32 candidate documents
- **Teacher targets:** listwise scores from `jinaai/jina-reranker-v3.5`

## How the hard negatives were mined

1. **Encode.** Queries and source corpora were encoded with LateOn-unsupervised into 128-dimensional
   token embeddings. This is the multi-vector model used for mining; it is the same
   late-interaction family as the final GLINT retriever.
2. **Retrieve by MaxSim.** Corpus shards were indexed with WARP and searched with the token-level
   MaxSim operator. WARP makes deep late-interaction retrieval tractable at mining scale; it is
   not required to run the released model.
3. **Merge shard rankings.** Per-shard candidate lists were merged by their MaxSim scores to form
   a global candidate pool for each query.
4. **Remove obvious positives.** The annotated positive was excluded. The very top miner ranks
   were skipped before candidate selection to reduce unlabelled-positive leakage.
5. **Construct KD lists.** The retained MaxSim candidates and gold anchors were deduplicated and
   assembled into fixed 32-document candidate sets. A frozen Jina listwise reranker then scored
   each complete set jointly; those scores are aligned position-for-position with `document_ids`.

The repository packages the prepared artifacts used by GLINT, not a re-ranker or an inference
index. WARP is therefore a data-construction dependency only; GLINT-base can be retrieved with a
standard late-interaction backend such as PyLate/PLAID.

## Contents

- `sft_bica/`: prepared SFT rows and metadata. The sources are MS MARCO, Natural Questions,
  HotpotQA, FEVER, FiQA, SQuAD v2, TriviaQA, and the BiCA citation-traversal augmentation.
- `kd_7src/`: the seven-source KD mixture in PyLate parquet layout (`train`, `queries`, and
  `documents`).
- `kd_7src_jina_scores/`: raw Jina listwise teacher-score shards keyed by query ID.

## Format

The KD mixture follows the PyLate layout:

| path | fields | notes |
|---|---|---|
| `train/train.parquet` | `query_id`, `document_ids`, `scores` | one row per query; `document_ids` has exactly 32 entries |
| `queries/train.parquet` | `query_id`, `text` | query text keyed by query ID |
| `documents/train.parquet` | `document_id`, `text` | candidate document text keyed by document ID |
| `kd_7src_jina_scores/*.parquet` | `query_id`, `scores` | 32 Jina scores in exactly the `document_ids` order |

## Composition

| subset | queries |
|---|---:|
| msmarco | 502,939 |
| nq | 152,132 |
| squadv2 | 130,217 |
| fever | 109,810 |
| hotpotqa | 85,000 |
| trivia | 60,413 |
| fiqa | 5,498 |
| **total** | **1,046,009** |

## Usage

```python
from datasets import load_dataset

# Load the prepared KD rows and their query/document tables.
train = load_dataset("chungimungi/GLINT-data", "kd_7src", split="train")
```