guidelabs/fineweb-atlas

FineWeb Atlas (v0.1)

FineWeb Atlas annotates 14.9 million FineWeb documents (95.5M chunks, 10.2B tokens) with 16,790 human-readable concepts spanning entities, topics, tones, and document types. Each chunk receives ~15 concept labels on average. The release includes chunk- and document-level annotations, a concept metadata table with prevalence stats, a reverse index for concept-first retrieval, and a packed cooccurrence matrix.

For background on how the atlas was built, see the companion blog post: The FineWeb Concept Atlas.

Quick start

from datasets import load_dataset

repo_id = "guidelabs/fineweb-atlas"

# Concept metadata and prevalence stats (small enough to load fully)
concepts = load_dataset(repo_id, "concepts", split="train").to_pandas()
concept_name_by_id = dict(zip(concepts["concept_id"], concepts["name"]))

# One row per document (streaming recommended for the full dataset)
documents = load_dataset(repo_id, "documents", split="train", streaming=True)
first_doc = next(iter(documents))

# Original chunk-level rows
chunks = load_dataset(repo_id, "chunks", split="train", streaming=True)
first_chunk = next(iter(chunks))

Resolve concept IDs to human-readable names:

def names_for_ids(ids, lookup):
    return [lookup[i] for i in ids]

# Example: show content concept names for the first chunk
print(names_for_ids(first_chunk["content_ids"], concept_name_by_id))

Load the full documents config into a pandas DataFrame locally. Prefer streaming=True unless you have ample RAM; the full chunks config is heavier.

from datasets import load_dataset

repo_id = "guidelabs/fineweb-atlas"
docs_df = load_dataset(repo_id, "documents", split="train").to_pandas()

Loading the cooccurrence matrix

The cooccurrence matrix is stored as a numpy file and is not loadable via load_dataset. Download it directly from the dataset repository:

import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="guidelabs/fineweb-atlas",
    filename="fineweb-concept-cooccurrence-matrix/fineweb-atlas-cooccurrence-upper-uint32.npy",
    repo_type="dataset",
)
packed = np.load(path, mmap_mode="r")

# Look up cooccurrence count for concepts i and j (where i <= j)
n = 16790
def cooc(i, j):
    if i > j:
        i, j = j, i
    return int(packed[i * n - (i * (i - 1)) // 2 + (j - i)])

# Example: how often do concepts 0 and 1 co-occur?
print(cooc(0, 1))

Available configs

Config	load_dataset call	Rows	Description
concepts (default)	load_dataset("guidelabs/fineweb-atlas", "concepts")	16,790	Concept metadata table with name, description, type, taxonomy, and prevalence stats.
documents	load_dataset("guidelabs/fineweb-atlas", "documents")	14,868,862	One row per document with full text, chunk character offsets, and document-level union concept lists by type.
chunks	load_dataset("guidelabs/fineweb-atlas", "chunks")	95,486,049	One row per chunk with chunk text, token span/count, quality status, and concept ID lists by type.
field_guide	load_dataset("guidelabs/fineweb-atlas", "field_guide")	1,406,432,869	One row per (concept_id, doc_int_id, chunk_id) assignment for concept-first retrieval.

The cooccurrence matrix is a numpy file and must be loaded separately (see Loading the cooccurrence matrix above).

All configs support streaming=True, which is recommended for the larger configs (documents, chunks, field_guide).

Which config should I use?

• Start with concepts if you want to browse the concept inventory, descriptions, prevalence, or taxonomy.
• Use documents if you need full-document text plus document-level union labels.
• Use chunks if you need per-chunk labels, token spans, or chunk-local context.
• Use field_guide if you need reverse lookup from a concept to matching chunks.

Shared IDs and joins

• concept_id is shared across all artifacts and is contiguous over 0..16789 (16790 concepts total).
• doc_int_id is the shared document key across documents, chunks, and field_guide.
• chunk_id identifies a chunk within a document and joins chunks to field_guide when paired with doc_int_id.
• document_text in documents is reconstructed by directly concatenating the ordered chunk texts. The chunker preserves inter-chunk whitespace as a prefix on each subsequent chunk, so direct concatenation is lossless.

Schemas

documents

One row per document.

Column	Type	Description
document_text	string	Full document text (direct concatenation of chunk texts, identical to the original FineWeb source).
doc_int_id	int32	Document key, shared across all configs.
chunk_count	int32	Number of chunks in this document.
document_token_count	int32	Total tokens across all chunks.
chunk_char_starts	list<int32>	Character offset where each chunk begins in document_text. Chunk i spans document_text[chunk_char_starts[i]:chunk_char_starts[i+1]] (last chunk runs to end of string).
chunk_token_starts	list<int32>	Token offset where each chunk begins in the document's token sequence.
chunk_token_counts	list<int32>	Number of tokens in each chunk.
has_long_chunk	bool	Whether any chunk exceeded the 128-token target.
has_segmentation_error	bool	Whether sentence packing failed (entire document is one chunk).
content_ids	list<int64>	Union of content concept IDs across all chunks (sorted, deduplicated).
tone_ids	list<int64>	Union of tone concept IDs across all chunks (sorted, deduplicated).
document_ids	list<int64>	Union of document-type concept IDs across all chunks (sorted, deduplicated).
entity_ids	list<int64>	Union of entity concept IDs across all chunks (sorted, deduplicated).

chunks

One row per chunk.

Column	Type	Description
doc_int_id	int32	Parent document key.
chunk_id	int16	Chunk index within the document (0-based).
chunk_text	string	Text content of this chunk.
chunk_token_start	int32	Inclusive token offset in the document.
chunk_token_end	int32	Exclusive token offset in the document.
chunk_token_count	int32	Number of tokens (chunk_token_end - chunk_token_start).
chunk_status	string	Quality status: ok, long_chunk, or segmentation_error.
tone_ids	list<int64>	Tone concept IDs assigned to this chunk.
entity_ids	list<int64>	Entity concept IDs assigned to this chunk.
content_ids	list<int64>	Content/topic concept IDs assigned to this chunk.
document_ids	list<int64>	Document-type concept IDs assigned to this chunk.

field_guide

One row per (concept, chunk) assignment. Useful for "which chunks mention concept X?" queries.

Column	Type	Description
concept_id	int32	Concept identifier (0..16789).
doc_int_id	int32	Document key.
chunk_id	int16	Chunk index within the document.

concepts

One row per concept.

Column	Type	Description
concept_id	uint32	Concept identifier (0..16789).
concept_type	string	One of entity, tone, content, document.
name	string	Human-readable concept name.
description	string	Short description of the concept.
taxonomy_lcc_path_primary	string	Top-level Library of Congress Classification class (single letter, e.g. H, T, Q) or "None".
chunk_count	int64	Number of chunks this concept appears in.
chunk_prevalence	float64	Fraction of all chunks containing this concept.

Cooccurrence matrix

Packed upper-triangular cooccurrence counts (with diagonal) over concept_id.

File	Description
fineweb-atlas-cooccurrence-upper-uint32.npy	Numpy array, dtype uint32, length 140,960,445 (= n*(n+1)/2, n=16790).
fineweb-atlas-cooccurrence-upper-uint32.json	Metadata (dtype, dimensions, max value).

Index mapping for 0 <= i <= j < n:

idx = i * n - (i * (i - 1)) // 2 + (j - i)
value = packed[idx]  # equals C[i, j]

The diagonal entry C[i, i] is the number of chunks containing concept i, which is also available directly as chunk_count in the concepts config.

Dataset statistics

Metric	Value
Documents	14,868,862
Chunks	95,486,049
Total tokens	10,183,028,973
Concepts	16,790
Reverse-index rows	1,406,432,869
Avg. labels per chunk	14.73
Avg. chunks per document	6.42
Avg. tokens per document	684.86
Median chunks per document	4

Labels per chunk by type:

Type	Avg. per chunk	Share of all labels	Concept count
Tone	6.90	46.8%	587
Content	5.68	38.6%	12,786
Document	1.52	10.3%	31
Entity	0.63	4.3%	3,386

Chunk quality: 97.5% ok, 2.5% long_chunk, 8 segmentation_error.

• ok: chunk is within the 128-token target.
• long_chunk: the sentence packer could not split the chunk further without breaking a sentence. These chunks exceed 128 tokens (median 171, max 54,408). Common causes are long unbroken paragraphs or code blocks.
• segmentation_error: sentence packing failed entirely and the whole document was placed into a single unsplit chunk. Only 8 documents (out of 14.9M) are affected — mostly very large spam pages or documents with unusual Unicode that confused the sentence splitter. These documents have concept annotations, but for the larger ones the labels may only reflect the beginning of the text due to annotation model context limits.

How it was built

FineWeb Atlas was produced using the ATLAS pipeline:

1. Chunking. FineWeb documents were split into chunks of ~128 tokens.
2. LLM annotation. Each chunk was annotated with structured concept tags by a language model, producing raw labels across four types: content topics, entities, tones, and document types.
3. Concept consolidation. Raw tags were embedded, clustered, and deduplicated into a canonical concept library of 16,790 concepts. Each concept received a human-readable name, description, and Library of Congress Classification path.
4. Scalable prediction. A trained set-prediction model applied the consolidated concept library to all 95.5M chunks.
5. Quality filtering. An LLM judge scored concept assignments, and a filtering policy determined final keep/drop decisions per concept-chunk pair.

For full details, see the blog post.

Common concepts: consider filtering

This release intentionally keeps very common labels, especially among tone and document concepts. They are useful for completeness, retrieval, and descriptive analysis, but they can dominate cooccurrence counts and downstream features if you are trying to surface more specific concepts.

In this release:

• 21 tone labels appear in at least 7% of chunks.
• 8 document labels appear in at least 7% of chunks.
• The most common overall labels are:
  • matter-of-fact (83.6%)
  • Informational (79.8%)
  • Brief and clear (69.0%)
  • factual (51.5%)
  • neutral (34.5%)
• The most common document label is Short announcement/bulletin (29.7%).

If you are training a downstream model, building sparse features, or trying to highlight more discriminative concepts, a reasonable first pass is to drop concepts above a prevalence threshold:

from datasets import load_dataset

repo_id = "guidelabs/fineweb-atlas"
concepts = load_dataset(repo_id, "concepts", split="train").to_pandas()
common_ids = set(concepts.loc[concepts["chunk_prevalence"] >= 0.07, "concept_id"])

def drop_common(ids):
    return [cid for cid in ids if cid not in common_ids]

The right threshold depends on the use case; 7% is a simple starting point.

Taxonomy distribution

Concepts are classified using the Library of Congress Classification system. The top-level class distribution:

Value	Domain	Concepts	Share
H	Social sciences	3,895	23.2%
T	Technology	1,959	11.7%
G	Geography, anthropology, recreation	1,792	10.7%
B	Philosophy, psychology, religion	1,332	7.9%
R	Medicine	1,076	6.4%
N	Fine arts	1,051	6.3%
Q	Science	856	5.1%
None	No primary class assigned	813	4.8%
J	Political science	682	4.1%
K	Law	539	3.2%
D	World history	496	3.0%
L	Education	491	2.9%
P	Language and literature	476	2.8%
E	History of the USA	314	1.9%
S	Agriculture	247	1.5%
C	Auxiliary sciences of history	195	1.2%
A	General works, reference	190	1.1%
Z	Bibliography, library science	144	0.9%
U	Military science	119	0.7%
F	History of the Americas	108	0.6%
V	Naval science	15	0.1%

These are the complete set of values for taxonomy_lcc_path_primary in the concepts config.

Relationship to FineWeb

This dataset annotates the official 10B-token subsample of FineWeb (sample/10BT), covering all 14.9M documents in that subsample. doc_int_id is a 1-based integer assigned sequentially during chunking and is stable within this release. It is not a FineWeb row index, but document_text in the documents config is identical to the original FineWeb source text and can be used for matching.

Notes and caveats

• Machine-generated labels. Concept assignments include noise, especially for rarer concepts.
• Source text quality. Some chunk text is noisy or garbled from the original web scrape; concept quality depends on chunk quality.
• No PII filtering. This dataset inherits the personal information characteristics of FineWeb. Web-scraped text may contain names, emails, or other personal information.
• English only. Annotations were produced for English text. Non-English passages that appear in FineWeb may have unreliable labels.
• documents is derived. The documents config is built by concatenating chunks. The text is identical to the FineWeb source, but for per-chunk detail (individual chunk statuses, per-chunk concept lists) use the chunks config.

Versioning

This is v0.1, the initial research release. The concept inventory, annotation model, and filtering policy may be revised in future versions. Breaking changes (e.g., concept ID renumbering) will increment the version.

Suggested citation

If you use this dataset, please cite:

@misc{monson_fineweb_concept_atlas_2026,
  author       = {Nathaniel Monson},
  title        = {The FineWeb Concept Atlas},
  year         = {2026},
  howpublished = {\url{https://www.guidelabs.ai/post/the-fineweb-concept-atlas/}},
  note         = {Guide Labs}
}