---
license: apache-2.0
task_categories: [text-retrieval]
tags: [colbert, late-interaction, multi-vector, embeddings, medical]
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
dataset_info:
  features:
    - name: id
      dtype: string
    - name: text
      dtype: string
    - name: n_tokens
      dtype: int32
    - name: embedding
      sequence: float16
  splits:
    - name: train
      num_examples: 200
---

# MIRIAD 200, encoded with mLateOn-medical

Multi-vector (ColBERT-style) embeddings for
[`tomaarsen/miriad-benchmark-200k`](https://huggingface.co/datasets/tomaarsen/miriad-benchmark-200k),
produced with [`multi-vector-encoder/mLateOn-medical`](https://huggingface.co/multi-vector-encoder/mLateOn-medical).

| | |
|---|---|
| passages | 200 |
| token vectors | 176,014 |
| mean vectors / passage | 880.07 |
| dim | 128 |
| stored dtype | float16 |
| embeddings size | 0.05 GB |
| raw text encoded | 1 MB |

The embeddings are **49x larger than the text
they came from**, which is why late-interaction retrieval needs quantization or pooling.

## Usage

`embedding` is flattened; reshape it to recover one vector per token.

```python
import numpy as np
from datasets import load_dataset

ds = load_dataset("KShivendu/miriad-mlateon-colbert-smoke", split="train")
row = ds[0]
vecs = np.asarray(row["embedding"], dtype=np.float16).reshape(row["n_tokens"], 128)
```

MaxSim, the late-interaction score:

```python
score = (query_vecs @ doc_vecs.T).max(axis=1).sum()
```

Queries, qrels and the raw layout live alongside in `queries.npy`, `queries_offs.npy`
and `qrels.json`.