FineWeb-Edu (Lance Format)

A Lance-formatted version of FineWeb-Edu — over 1.5 billion educational web passages with cleaned text, source metadata, language detection signals, and 384-dim text embeddings — available directly from the Hub at hf://datasets/lance-format/fineweb-edu/data/train.lance.

Key features

• Cleaned passage text in the text column with the source url and title carried alongside.
• Language detection signals (language, language_probability) for filtered subsets.
• Pre-computed 384-dim text embeddings in text_embedding, ready for ANN search once an index is built locally.
• One columnar dataset — scan metadata cheaply, project just the columns each query needs, defer the heavy text and text_embedding reads to the rows that matter.

No pre-built indices on the Hub copy yet. At 1.5 B+ rows the on-disk indices are too large to ship comfortably alongside the data on the Hub. The Search, Curate, Evolve, and Train sections below describe the same APIs you'd use against a fully indexed dataset, but vector and full-text examples assume a local copy with IVF_PQ and INVERTED indices built once after download. See the Materialize-a-subset section at the end for a focused-subset workflow that makes indexing tractable.

Splits

train.lance

Schema

Column	Type	Notes
text	string	Cleaned passage body
title	string	Page or article title when available
url	string	Canonical source URL
language	string	Detected language code (e.g., en)
language_probability	float32	Confidence of the language detector
text_embedding	fixed_size_list<float32, 384>	Passage embedding for retrieval
FineWeb-Edu quality metadata	—	Heuristic scores and length statistics carried over from the upstream corpus

Pre-built indices

None bundled at present. Build the recommended indices on a local copy:

import lancedb

db = lancedb.connect("./fineweb-edu/data")
tbl = db.open_table("train")

tbl.create_index(
    metric="cosine",
    vector_column_name="text_embedding",
    index_type="IVF_PQ",
    num_partitions=2048,
    num_sub_vectors=96,
)
tbl.create_fts_index("text", replace=True)

Both indices live next to the data, so subsequent queries against the same local path pick them up automatically.

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/fineweb-edu", split="train", streaming=True)
for row in hf_ds.take(3):
    print(row["title"] or row["url"])

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, Versioning, and Materialize-a-subset sections below.

import lancedb

db = lancedb.connect("hf://datasets/lance-format/fineweb-edu/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, the list of pre-built indices.

import lance

ds = lance.dataset("hf://datasets/lance-format/fineweb-edu/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 at 1.5 B+ rows random access and any kind of search are dramatically faster against a local copy, and ANN / FTS require local indices anyway:

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

Then point Lance or LanceDB at ./fineweb-edu/data. For most workflows, the Materialize-a-subset section is a better starting point than downloading the full 1.5 B-row corpus.

Search

Once an IVF_PQ index exists on text_embedding, dense retrieval is a single call. In production you would encode a query string through the same 384-dim text encoder used at ingest and pass the resulting vector to tbl.search(...). The example below uses the embedding from row 42 as a runnable stand-in.

import lancedb

db = lancedb.connect("./fineweb-edu/data")  # local copy with the indices from the section above
tbl = db.open_table("train")

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

hits = (
    tbl.search(seed["text_embedding"])
    .metric("cosine")
    .where("language = 'en' AND language_probability > 0.9", prefilter=True)
    .select(["title", "url", "text"])
    .limit(10)
    .to_list()
)
for r in hits:
    print(f"{r['url']}\n  {(r['title'] or '')[:80]}")

The result set carries only the projected columns. The text_embedding vector is never read on the result side, and the text body is fetched only for the ten passages that actually came back, keeping the working set small even though the corpus is enormous.

Because the recommended setup also builds an INVERTED index on text, 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 must literally appear in the passage but the dense side still does most of the ranking.

hybrid_hits = (
    tbl.search(query_type="hybrid")
    .vector(seed["text_embedding"])
    .text("quantum computing")
    .where("language = 'en'", prefilter=True)
    .select(["title", "url", "text"])
    .limit(10)
    .to_list()
)
for r in hybrid_hits:
    print(f"{r['url']}\n  {(r['title'] or '')[:80]}")

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

Curate

A typical curation pass over a web corpus starts with a metadata filter — pick high-confidence English, drop short or low-quality fragments, restrict to a domain — before any text gets read. Stacking predicates inside a single filtered scan keeps the result small and explicit, and the bounded .limit(1000) makes it cheap to inspect.

import lancedb

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

candidates = (
    tbl.search()
    .where(
        "language = 'en' "
        "AND language_probability > 0.95 "
        "AND length(text) >= 1000",
        prefilter=True,
    )
    .select(["url", "title", "language_probability"])
    .limit(1000)
    .to_list()
)
print(f"{len(candidates)} candidates; first url: {candidates[0]['url']}")

The result is a plain list of dictionaries, ready to inspect, persist as a manifest of URLs, or hand to the Materialize-a-subset section below for export to a writable local copy. Neither the text body nor the text_embedding vector is read by this scan, so a 1000-row curation pass against the Hub moves only kilobytes of metadata even though the underlying table is in the billions.

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 a text_length and a long_passage 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 a larger slice first.

import lancedb

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

tbl.add_columns({
    "text_length": "length(text)",
    "long_passage": "length(text) >= 1000",
})

If the values you want to attach already live in another table (offline labels, topic classifications, alternate embeddings from a stronger model), merge them in by joining on url:

import pyarrow as pa

labels = pa.table({
    "url": pa.array(["https://example.com/a", "https://example.com/b"]),
    "topic": pa.array(["math", "history"]),
})
tbl.merge(labels, on="url")

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 different embedding model over the text), 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 language-model pretraining the natural projection is just the text column; for a retrieval probe or a reranker on top of frozen features, project the precomputed embedding instead.

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

db = lancedb.connect("./fineweb-edu/data")
tbl = db.open_table("train")

train_ds = Permutation.identity(tbl).select_columns(["text"])
loader = DataLoader(train_ds, batch_size=64, shuffle=True, num_workers=8)

for batch in loader:
    # batch carries only the projected columns; tokenize, forward, backward...
    ...

Switching feature sets is a configuration change: passing ["text_embedding"] to select_columns(...) on the next run reads only the 384-d vectors and skips the text body entirely, which is the right shape for training a lightweight retrieval head on cached embeddings. Columns added in Evolve cost nothing per batch until they are explicitly projected.

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/fineweb-edu/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("./fineweb-edu/data")
local_tbl = local_db.open_table("train")
local_tbl.tags.create("english-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="english-v1")
tbl_v5 = db.open_table("train", version=5)

Pinning supports two workflows. A retrieval system locked to english-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 passages, so changes in metrics reflect model changes rather than data drift. Neither workflow needs shadow copies or external manifest tracking.

Materialize a subset

At 1.5 B+ rows, very few workflows want the full corpus on local disk. The practical entry point 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. The result is a writable LanceDB database scoped to the rows that actually matter for the downstream task, sized to index and iterate cheaply.

import lancedb

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

batches = (
    remote_tbl.search()
    .where(
        "language = 'en' "
        "AND language_probability > 0.95 "
        "AND length(text) >= 1000"
    )
    .select(["url", "title", "text", "language", "language_probability", "text_embedding"])
    .to_batches()
)

local_db = lancedb.connect("./fineweb-edu-en")
local_db.create_table("train", batches)

The resulting ./fineweb-edu-en is a first-class LanceDB database. Build the recommended indices on it once (the same create_index / create_fts_index calls shown in the Pre-built indices section, pointed at the local path), and every snippet in the Search, Evolve, Train, and Versioning sections above works against it by swapping hf://datasets/lance-format/fineweb-edu/data for ./fineweb-edu-en.

Source & license

Converted from HuggingFaceFW/fineweb-edu. FineWeb-Edu is distributed under ODC-BY 1.0; individual document content remains subject to the rights of the original publishers. Review the upstream dataset card before downstream use.

Citation

@misc{lozhkov2024finewebedu,
  title  = {FineWeb-Edu: the Finest Collection of Educational Content the Web Has to Offer},
  author = {Lozhkov, Anton and Ben Allal, Loubna and von Werra, Leandro and Wolf, Thomas},
  year   = {2024},
  url    = {https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu}
}