Flickr30k

A Lance-formatted version of Flickr30k, redistributed via lmms-lab/flickr30k. Each row is one image with 5 human-written captions, a cosine-normalized CLIP image embedding, and a cosine-normalized CLIP text embedding of the canonical caption — all…

View on Hugging Face

Source dataset card and downloadable files for lance-format/flickr30k-lance.

A Lance-formatted version of Flickr30k, redistributed via lmms-lab/flickr30k. Each row is one image with 5 human-written captions, a cosine-normalized CLIP image embedding, and a cosine-normalized CLIP text embedding of the canonical caption — all stored inline and available directly from the Hub at hf://datasets/lance-format/flickr30k-lance/data.

Key features

• Inline JPEG bytes in the image column — no sidecar files, no image folders.
• Paired CLIP embeddings in the same row — image_emb and text_emb (ViT-B/32, 512-dim, cosine-normalized) — so cross-modal retrieval is one indexed lookup.
• All 5 raw captions kept in captions alongside a caption canonical string used for full-text search.
• Pre-built ANN, FTS, and scalar indices covering both embedding columns, the canonical caption, and image_id.

Splits

Split	Rows	Notes
train.lance	31,783	All Flickr30k images; the lmms-lab/flickr30k redistribution merges the original train/val/test labels into a single split

Schema

Column	Type	Notes
id	int64	Row index within split (natural join key)
image	large_binary	Inline JPEG bytes
image_id	string	Original Flickr image id
filename	string?	Original filename (e.g. 1000092795.jpg)
captions	list<string>	All 5 captions for the image
caption	string	First caption — canonical text used for FTS
image_emb	fixed_size_list<float32, 512>	CLIP image embedding (cosine-normalized)
text_emb	fixed_size_list<float32, 512>	CLIP text embedding of the canonical caption

Pre-built indices

• IVF_PQ on image_emb — image-side vector search (cosine)
• IVF_PQ on text_emb — text-side vector search (cosine)
• INVERTED (FTS) on caption — keyword and hybrid search
• BTREE on image_id — fast lookup by Flickr image id

Why Lance?

1. Blazing Fast Random Access: Optimized for fetching scattered rows, making it ideal for random sampling, real-time ML serving, and interactive applications without performance degradation.
2. Native Multimodal Support: Store text, embeddings, and other data types together in a single file. Large binary objects are loaded lazily, and vectors are optimized for fast similarity search.
3. Native Index Support: Lance comes with fast, on-disk, scalable vector and FTS indexes that sit right alongside the dataset on the Hub, so you can share not only your data but also your embeddings and indexes without your users needing to recompute them.
4. Efficient Data Evolution: Add new columns and backfill data without rewriting the entire dataset. This is perfect for evolving ML features, adding new embeddings, or introducing moderation tags over time.
5. Versatile Querying: Supports combining vector similarity search, full-text search, and SQL-style filtering in a single query, accelerated by on-disk indexes.
6. Data Versioning: Every mutation commits a new version; previous versions remain intact on disk. Tags pin a snapshot by name, so retrieval systems and training runs can reproduce against an exact slice of history.

Load with datasets.load_dataset

You can load Lance datasets via the standard HuggingFace datasets interface, suitable when your pipeline already speaks Dataset / IterableDataset or you want a quick streaming sample.

import datasets

hf_ds = datasets.load_dataset("lance-format/flickr30k-lance", split="train", streaming=True)
for row in hf_ds.take(3):
    print(row["caption"])

Load with LanceDB

LanceDB is the embedded retrieval library built on top of the Lance format (docs), and is the interface most users interact with. It wraps the dataset as a queryable table with search and filter builders, and is the entry point used by the Search, Curate, Evolve, Train, Versioning, and Materialize-a-subset sections below.

import lancedb

db = lancedb.connect("hf://datasets/lance-format/flickr30k-lance/data")
tbl = db.open_table("train")
print(len(tbl))

Load with Lance

pylance is the Python binding for the Lance format and works directly with the format’s lower-level APIs. Reach for it when you want to inspect dataset internals — schema, scanner, fragments, and the list of pre-built indices.

import lance

ds = lance.dataset("hf://datasets/lance-format/flickr30k-lance/data/train.lance")
print(ds.count_rows(), ds.schema.names)
print(ds.list_indices())

Tip — for production use, download locally first. Streaming from the Hub works for exploration, but heavy random access and ANN search are far faster against a local copy:

hf download lance-format/flickr30k-lance --repo-type dataset --local-dir ./flickr30k-lance

Then point Lance or LanceDB at ./flickr30k-lance/data.

Search

The bundled IVF_PQ index on image_emb makes cross-modal text→image retrieval a single call: encode a text query with the same CLIP model used at ingest (ViT-B/32, cosine-normalized), then pass the resulting 512-d vector to tbl.search(...) and target image_emb. The example below uses the text_emb already stored in row 42 as a runnable stand-in for “the CLIP encoding of a caption”, so the snippet works without any model loaded.

import lancedb

db = lancedb.connect("hf://datasets/lance-format/flickr30k-lance/data")
tbl = db.open_table("train")

seed = (
    tbl.search()
    .select(["text_emb", "caption"])
    .limit(1)
    .offset(42)
    .to_list()[0]
)

hits = (
    tbl.search(seed["text_emb"], vector_column_name="image_emb")
    .metric("cosine")
    .select(["image_id", "caption"])
    .limit(10)
    .to_list()
)
print("query caption:", seed["caption"])
for r in hits:
    print(f"  {r['image_id']:>12}  {r['caption'][:70]}")

Because OpenAI-style CLIP embeddings are normalized, cosine is the right metric and the first hit will typically be the source image itself — a useful sanity check. Swap vector_column_name="image_emb" for text_emb to do text→text retrieval against the canonical captions instead.

Because the dataset also ships an INVERTED index on caption, the same query can be issued as a hybrid search that combines the dense vector with a keyword query. LanceDB merges the two result lists and reranks them in a single call, which is useful when a phrase like “dog playing in the snow” must literally appear in the caption but you still want CLIP to do the heavy lifting on visual similarity.

hybrid_hits = (
    tbl.search(query_type="hybrid", vector_column_name="image_emb")
    .vector(seed["text_emb"])
    .text("dog playing in the snow")
    .select(["image_id", "caption"])
    .limit(10)
    .to_list()
)
for r in hybrid_hits:
    print(f"  {r['image_id']:>12}  {r['caption'][:70]}")

Tune metric, nprobes, and refine_factor on the vector side to trade recall against latency.

Curate

A typical curation pass for a captioning or contrastive-training workflow combines a content filter on the captions with a structural filter on the row. Stacking both inside a single filtered scan keeps the result small and explicit, and the bounded .limit(500) makes it cheap to inspect before committing the subset to anything downstream.

import lancedb

db = lancedb.connect("hf://datasets/lance-format/flickr30k-lance/data")
tbl = db.open_table("train")

candidates = (
    tbl.search("surfer OR surfboard OR wave")
    .where("array_length(captions) = 5", prefilter=True)
    .select(["image_id", "caption", "captions"])
    .limit(500)
    .to_list()
)
print(f"{len(candidates)} candidates; first caption: {candidates[0]['caption'][:80]}")

The result is a plain list of dictionaries, ready to inspect, persist as a manifest of image_ids, or feed into the Evolve and Train workflows below. The image column is never read, so the network traffic for a 500-row candidate scan is dominated by caption text rather than JPEG bytes.

Evolve

Lance stores each column independently, so a new column can be appended without rewriting the existing data. The lightest form is a SQL expression: derive the new column from columns that already exist, and Lance computes it once and persists it. The example below adds num_captions and a long_caption flag, either of which can then be used directly in where clauses without recomputing the predicate on every query.

Note: Mutations require a local copy of the dataset, since the Hub mount is read-only. See the Materialize-a-subset section at the end of this card for a streaming pattern that downloads only the rows and columns you need, or use hf download to pull the full split first.

import lancedb

db = lancedb.connect("./flickr30k-lance/data")  # local copy required for writes
tbl = db.open_table("train")

tbl.add_columns({
    "num_captions": "array_length(captions)",
    "long_caption": "length(caption) >= 80",
})

If the values you want to attach already live in another table (offline labels, classifier predictions, an aesthetic or NSFW score, a second-pass caption from a different model), merge them in by joining on image_id:

import pyarrow as pa

labels = pa.table({
    "image_id": pa.array(["1000092795", "10002456"]),
    "scene_label": pa.array(["outdoor", "indoor"]),
})
tbl.merge(labels, on="image_id")

The original columns and indices are untouched, so existing code that does not reference the new columns continues to work unchanged. New columns become visible to every reader as soon as the operation commits. For column values that require a Python computation (e.g., running a second CLIP variant over the image bytes), Lance provides a batch-UDF API — see the Lance data evolution docs.

Train

Projection lets a training loop read only the columns each step actually needs. LanceDB tables expose this through Permutation.identity(tbl).select_columns([...]), which plugs straight into the standard torch.utils.data.DataLoader so prefetching, shuffling, and batching behave as in any PyTorch pipeline. For a CLIP-style contrastive run, project the JPEG bytes and a sampled caption; for a reranker or probe on top of frozen features, project the precomputed embeddings instead.

import lancedb
from lancedb.permutation import Permutation
from torch.utils.data import DataLoader

db = lancedb.connect("hf://datasets/lance-format/flickr30k-lance/data")
tbl = db.open_table("train")

train_ds = Permutation.identity(tbl).select_columns(["image", "caption"])
loader = DataLoader(train_ds, batch_size=128, shuffle=True, num_workers=4)

for batch in loader:
    # batch carries only the projected columns; decode the JPEG bytes,
    # tokenize the captions, encode, contrastive loss...
    ...

Switching feature sets is a configuration change: passing ["image_emb", "text_emb"] to select_columns(...) on the next run skips JPEG decoding entirely and reads only the cached 512-d vectors, which is the right shape for training a lightweight reranker or a linear probe.

Versioning

Every mutation to a Lance dataset, whether it adds a column, merges labels, or builds an index, commits a new version. Previous versions remain intact on disk. You can list versions and inspect the history directly from the Hub copy; creating new tags requires a local copy since tags are writes.

import lancedb

db = lancedb.connect("hf://datasets/lance-format/flickr30k-lance/data")
tbl = db.open_table("train")

print("Current version:", tbl.version)
print("History:", tbl.list_versions())
print("Tags:", tbl.tags.list())

Once you have a local copy, tag a version for reproducibility:

local_db = lancedb.connect("./flickr30k-lance/data")
local_tbl = local_db.open_table("train")
local_tbl.tags.create("clip-vitb32-v1", local_tbl.version)

A tagged version can be opened by name, or any version reopened by its number, against either the Hub copy or a local one:

tbl_v1 = db.open_table("train", version="clip-vitb32-v1")
tbl_v5 = db.open_table("train", version=5)

Pinning supports two workflows. A retrieval system locked to clip-vitb32-v1 keeps returning stable results while the dataset evolves in parallel — newly added embeddings or labels do not change what the tag resolves to. A training experiment pinned to the same tag can be rerun later against the exact same images and captions, so changes in metrics reflect model changes rather than data drift. Neither workflow needs shadow copies or external manifest tracking.

Materialize a subset

Reads from the Hub are lazy, so exploratory queries only transfer the columns and row groups they touch. Mutating operations (Evolve, tag creation) need a writable backing store, and a training loop benefits from a local copy with fast random access. Both can be served by a subset of the dataset rather than the full split. The pattern is to stream a filtered query through .to_batches() into a new local table; only the projected columns and matching row groups cross the wire, and the bytes never fully materialize in Python memory.

import lancedb

remote_db = lancedb.connect("hf://datasets/lance-format/flickr30k-lance/data")
remote_tbl = remote_db.open_table("train")

batches = (
    remote_tbl.search("surfer OR surfboard OR wave")
    .where("array_length(captions) = 5")
    .select(["image_id", "image", "caption", "captions", "image_emb", "text_emb"])
    .to_batches()
)

local_db = lancedb.connect("./flickr30k-surf-subset")
local_db.create_table("train", batches)

The resulting ./flickr30k-surf-subset is a first-class LanceDB database. Every snippet in the Evolve, Train, and Versioning sections above works against it by swapping hf://datasets/lance-format/flickr30k-lance/data for ./flickr30k-surf-subset.

Source & license

Converted from lmms-lab/flickr30k, which is itself a parquet redistribution of the original Flickr30k corpus. Original images come from Flickr; review the Flickr30k licensing terms before redistribution.

Citation

@article{young2014image,
  title={From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions},
  author={Young, Peter and Lai, Alice and Hodosh, Micah and Hockenmaier, Julia},
  journal={Transactions of the Association for Computational Linguistics},
  volume={2},
  pages={67--78},
  year={2014}
}