---
language:
- en
pretty_name: "SIFT1M-variant Dataset for (Filtered) Approximate Nearest Neighbor Search"
license: mit
task_categories:
- text-retrieval
tags:
- vector-search
- approximate-nearest-neighbor
- approximate-nearest-neighbor-search
- filtered-approximate-nearest-neighbor-search
- ann-benchmark
dataset_info:
  features:
  - name: id
    dtype: int64
  - name: label
    dtype: int64
  - name: range_value
    dtype: float64
size_categories:
- 1M<n<10M
---

# SIFT1M Dataset to Evaluate (Filtered) Approximate Nearest Neighbor Search — Normal Variant

This dataset is intended to benchmark Approximate Nearest Neighbor Search (ANNS) and
Filtered Approximate Nearest Neighbor Search (FANNS) algorithms. It is based on the
classic **SIFT1M** dataset (1,000,000 base vectors and 10,000 query vectors, each
128-dimensional). Because SIFT1M ships without any structured metadata, we augment it
with **synthetic attributes drawn from a normal (Gaussian) distribution**: every base
vector receives one categorical `label` and one continuous `range_value`. Query
attributes for three different filter types are sampled from the same normal distribution
as the base attributes. The ground truth for the up to `k=100` nearest neighbors was
computed for unfiltered ANNS and for filtered ANNS with the three different filter types.
Please note that if fewer than `100` database items match the filter, then the ground
truth can contain fewer than `k=100` entries. This variant is the Gaussian counterpart of
**sift1m-fanns-uniform**: the design is identical and only the sampling distribution
changes, so that the two can be compared directly to study the effect of attribute
distribution (homogeneous vs. heterogeneous selectivity).

## Filter Types

| Filter | Predicate | Description |
| ------ | --------- | ----------- |
| label  | `label == required_label` | Exact match on the categorical attribute. |
| range  | `range_low <= range_value <= range_high` | Containment in a fixed-width interval. |
| both   | `label` AND `range` | Conjunction of the two predicates above. |

## Parameters

| Parameter | Value | Meaning |
| --------- | ----- | ------- |
| N | 1,000,000 | Number of base (database) vectors |
| nq | 10,000 | Number of query vectors |
| vec_dim | 128 | Embedding dimension (from SIFT1M) |
| NUM_LABELS | 12 | `label ~ round(Normal(5.5, 2.0))` clipped to [0, 11] |
| LABEL_MEAN / LABEL_STD | 5.5 / 2.0 | Mean and std of the label histogram |
| range_value | `Normal(0.5, 0.15)` clipped to [0, 1] | Continuous base attribute |
| RANGE_MEAN / RANGE_STD | 0.5 / 0.15 | Mean and std of `range_value` and of query window centers |
| RANGE_WIDTH | 0.2 | Width of every query range window (`range_high - range_low`) |
| GT_K | 100 | Max number of ground-truth neighbors per query |
| RANDOM_SEED | 42 | Fixed seed; the dataset is bit-reproducible |

## Expected Selectivity

Under the normal distribution, the **average** selectivity is higher than in the uniform
variant, but more importantly the **per-query selectivity has high variance**
(heterogeneous): common labels and central range windows are far less selective than rare
labels and edge windows. A single benchmark therefore spans a wide range of difficulties,
which is closer to real-world skewed attributes.

| Filter | Avg. selectivity (approx.) | Per-query spread | Note |
| ------ | -------------------------- | ---------------- | ---- |
| label  | ~14%  | ~0.6% (label 0/11) → ~19% (label 5/6) | `Σ pₓ²` exceeds the uniform 8.33% |
| range  | ~30–40% | ~8% (edge window) → ~50% (central window) | Windows concentrate near the dense center |
| both   | ~5%   | <0.5% (rare label + edge window) → ~10% | Highest variance; a few queries may yield < 100 GT entries |

## Files and Description

| File | Description |
| ---- | ----------- |
| `database_vectors.fvecs` | 128-dimensional base vectors. One vector per database item. |
| `database_attributes.jsonl` | JSON objects with `{id, label, range_value}` for each item. One JSON object per database item. |
| `query_vectors.fvecs` | 128-dimensional query vectors. One vector per query. |
| `ground_truth.ivecs` | Ground truth for unfiltered nearest neighbor search. One vector per query. |
| `label_query_attributes.jsonl` | Query attributes for label (exact match) filtering. One JSON object per query. |
| `ground_truth_label.ivecs` | Ground truth for label-filtered NN search. One vector per query. |
| `range_query_attributes.jsonl` | Query attributes for range filtering. One JSON object per query. |
| `ground_truth_range.ivecs` | Ground truth for range-filtered NN search. One vector per query. |
| `label_and_range_query_attributes.jsonl` | Query attributes for the joint (label AND range) filter. One JSON object per query. |
| `ground_truth_label_and_range.ivecs` | Ground truth for joint-filtered NN search. One vector per query. |

## Formats

- `.fvecs`: Binary format for 32-bit floating point numbers (used for embedding vectors).
- `.ivecs`: Binary format for 32-bit signed integers (used for ground truth).
- `.jsonl`: Each line contains a JSON object (used for attributes).