---
license: apache-2.0
language:
  - en
pretty_name: HSEB MSMARCO benchmarking dataset
dataset_info:
  features:
  - name: text
    dtype: string
  - name: id
    dtype: int64
  - name: embedding
    sequence: float32
  - name: results_10_docs
    sequence: int64
  - name: results_10_scores
    sequence: float32
  - name: results_90_docs
    sequence: int64
  - name: results_90_scores
    sequence: float32
  - name: results_100_docs
    sequence: int64
  - name: results_100_scores
    sequence: float32
  - name: tag
    sequence: int64
configs:
  - config_name: "query-all-MiniLM-L6-v2-1K"
    data_files: "data/all-MiniLM-L6-v2/1K/queries.jsonl.gz"
    default: true
  - config_name: "corpus-all-MiniLM-L6-v2-1K"
    data_files: "data/all-MiniLM-L6-v2/1K/corpus.jsonl.gz"
  - config_name: "query-all-MiniLM-L6-v2-100K"
    data_files: "data/all-MiniLM-L6-v2/100K/queries.jsonl.gz"
  - config_name: "corpus-all-MiniLM-L6-v2-100K"
    data_files: "data/all-MiniLM-L6-v2/100K/corpus.jsonl.gz"
  - config_name: "query-all-MiniLM-L6-v2-1M"
    data_files: "data/all-MiniLM-L6-v2/1M/queries.jsonl.gz"
  - config_name: "corpus-all-MiniLM-L6-v2-1M"
    data_files: "data/all-MiniLM-L6-v2/1M/corpus.jsonl.gz"

  - config_name: "query-e5-base-v2-1K"
    data_files: "data/e5-base-v2/1K/queries.jsonl.gz"
  - config_name: "corpus-e5-base-v2-1K"
    data_files: "data/e5-base-v2/1K/corpus.jsonl.gz"
  - config_name: "query-e5-base-v2-100K"
    data_files: "data/e5-base-v2/100K/queries.jsonl.gz"
  - config_name: "corpus-e5-base-v2-100K"
    data_files: "data/e5-base-v2/100K/corpus.jsonl.gz"
  - config_name: "query-e5-base-v2-1M"
    data_files: "data/e5-base-v2/1M/queries.jsonl.gz"
  - config_name: "corpus-e5-base-v2-1M"
    data_files: "data/e5-base-v2/1M/corpus.jsonl.gz"

  - config_name: "query-Qwen3-Embedding-4B-1K"
    data_files: "data/Qwen3-Embedding-4B/1K/queries.jsonl.gz"
  - config_name: "corpus-Qwen3-Embedding-4B-1K"
    data_files: "data/Qwen3-Embedding-4B/1K/corpus.jsonl.gz"
  - config_name: "query-Qwen3-Embedding-4B-100K"
    data_files: "data/Qwen3-Embedding-4B/100K/queries.jsonl.gz"
  - config_name: "corpus-Qwen3-Embedding-4B-100K"
    data_files: "data/Qwen3-Embedding-4B/100K/corpus.jsonl.gz"
  - config_name: "query-Qwen3-Embedding-4B-1M"
    data_files: "data/Qwen3-Embedding-4B/1M/queries.jsonl.gz"
  - config_name: "corpus-Qwen3-Embedding-4B-1M"
    data_files: "data/Qwen3-Embedding-4B/1M/corpus.jsonl.gz"

---

# HSEB MSMARCO benchmarking dataset

This collection is based on [MSMARCO](TODO) dataset:

* Embedding models: 
    * 384 dims: [sentence-transformers/all-MiniLM-L6-v2](TODO)
    * 768 dims: [intfloat/e5-base-v2](TODO)
    * 2560 dims: [Qwen3-Embedding-4B](TODO)
* Splits:
    * `1K`: 1K documents, 10K queries
    * `100K`: 100K documents, 10K queries
    * `1M`: 1M documents, 10K queries
* Filter selectivity:
    * `10%` for high selectivity, `90%` for low selectivity, `100%` for no filters at all
    * each document has a tag based on sampled selectivity, so 10% of docs have a tag `10`, and 50% of docs have tag `50`
* Exact match results:
    * for each selectivity level for each query there are precomputed exact k-NN search results for top-100 documents.


## Dataset structure

The dataset can be loaded with the [Huggingface datasets](TODO) library:

```python
from datasets import load_dataset

query = load_dataset("hseb-benchmark/msmarco", "query-all-MiniLM-L6-v2-1M")
corpus = load_dataset("hseb-benchmark/msmarco", "corpus-all-MiniLM-L6-v2-1M")
```

where 2nd argument to `load_dataset` is a config name, consisting of the following parts:

```
<query|corpus>-<model>-<size>
```

1. `query` | `corpus` - which side of the dataset you're loading
2. `model`: embedding model, `all-MiniLM-L6-v2` (384 dims), `e5-base-v2` (768 dims) or `Qwen3-Embedding-4B` (2560 dims)
3. `size`: sample size, `1K`, `100K`, `1M`

## License

Apache 2.0