Datasets:

guidelabs
/

fineweb-atlas

concept_id int32 0 16.8k ⌀	doc_int_id int32 1 14.9M	chunk_id int16 0 1.57k
1,713	173	2
1,722	173	2
4,682	173	2
9,706	173	2
12,806	173	2
15,685	173	2
16,230	173	2
16,430	173	2
16,500	173	2
16,503	173	2
16,544	173	2
16,642	173	2
16,731	173	2
2,683	176	5
3,218	176	5
4,567	176	5
6,147	176	5
8,084	176	5
10,211	176	5
12,806	176	5
12,812	176	5
14,279	176	5
16,259	176	5
16,270	176	5
16,353	176	5
16,410	176	5
16,500	176	5
16,544	176	5
16,595	176	5
2,683	176	11
3,218	176	11
4,567	176	11
6,147	176	11
10,211	176	11
11,621	176	11
12,806	176	11
12,812	176	11
15,507	176	11
16,259	176	11
16,270	176	11
16,353	176	11
16,410	176	11
16,500	176	11
16,544	176	11
16,593	176	11
2,189	179	2
7,949	179	2
10,773	179	2
12,093	179	2
12,094	179	2
12,471	179	2
12,812	179	2
13,958	179	2
14,721	179	2
16,259	179	2
16,410	179	2
16,430	179	2
16,482	179	2
16,503	179	2
16,539	179	2
16,544	179	2
1,837	181	2
8,061	181	2
8,409	181	2
10,068	181	2
12,795	181	2
12,809	181	2
16,259	181	2
16,430	181	2
16,503	181	2
16,511	181	2
16,544	181	2
16,611	181	2
16,624	181	2
16,733	181	2
2,649	183	5
2,696	183	5
2,699	183	5
2,741	183	5
3,543	183	5
5,315	183	5
11,185	183	5
12,814	183	5
15,905	183	5
16,309	183	5
16,399	183	5
16,430	183	5
16,444	183	5
16,503	183	5
16,539	183	5
16,544	183	5
16,561	183	5
16,642	183	5
3,832	187	5
6,975	187	5
9,570	187	5
11,556	187	5
11,558	187	5
11,925	187	5
12,804	187	5

End of preview. Expand in Data Studio

FineWeb Atlas (v0)

This dataset repository contains machine-generated concept annotations over FineWeb chunks, plus a concept reverse index and a packed concept cooccurrence matrix.

What is in `v0/`

v0/fineweb-atlas-annotated/ (57 parquet shards)
v0/fineweb-atlas-annotated-reverse-index/ (57 parquet shards)
v0/fineweb-concept-cooccurrence-matrix/
- fineweb-atlas-cooccurrence-upper-uint32.npy
- fineweb-atlas-cooccurrence-upper-uint32.json
v0/fineweb-concept-atlas.parquet

Artifact guide

Artifact	Path(s)	What it is
`chunks`	`v0/fineweb-atlas-annotated/*.parquet`	One row per chunk with chunk text, token span/count, status, and concept ID lists by type (`content_ids`, `tone_ids`, `document_ids`, `entity_ids`).
`field_guide` (reverse index)	`v0/fineweb-atlas-annotated-reverse-index/*.parquet`	One row per `(concept_id, doc_int_id, chunk_id)` assignment for concept-first retrieval.
`concepts`	`v0/fineweb-concept-atlas.parquet`	Concept metadata table (`concept_id`, `concept_type`, `name`, `description`, `taxonomy_lcc_path_primary`).
`cooccurrence`	`v0/fineweb-concept-cooccurrence-matrix/*`	Packed upper-triangular concept cooccurrence matrix (`uint32`) plus metadata JSON.

Notes:

doc_int_id in chunks is the document key for grouping/joining chunk rows.

Core ID contract

concept_id is shared across all artifacts and is contiguous:

range: 0..16789
number of concepts: 16790

Artifact schemas

`fineweb-atlas-annotated` (chunk-level)

One row per chunk.

doc_int_id: int32
chunk_id: int16
chunk_text: string
chunk_token_start: int32
chunk_token_end: int32
chunk_token_count: int32
chunk_status: string
entity_ids: list<int64>
tone_ids: list<int64>
content_ids: list<int64>
document_ids: list<int64>

`fineweb-atlas-annotated-reverse-index`

One row per (concept_id, chunk) assignment.

concept_id: int32
doc_int_id: int32
chunk_id: int16

`fineweb-concept-atlas`

Concept table.

concept_id: uint32
concept_type: string (entity, tone, content, document)
name: string
description: string
taxonomy_lcc_path_primary: string

Chunk token span fields (examples):

chunk_token_start: inclusive token offset where the chunk begins in the document (example: 256)
chunk_token_end: exclusive token offset where the chunk ends in the document (example: 384)
chunk_token_count: number of tokens in the chunk; typically chunk_token_end - chunk_token_start (example: 128)
chunk_status: segmentation/quality status for the chunk (examples: ok, long_chunk, segmentation_error)

`fineweb-concept-cooccurrence-matrix`

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

matrix file: fineweb-atlas-cooccurrence-upper-uint32.npy
metadata: fineweb-atlas-cooccurrence-upper-uint32.json
dtype: uint32
length: 140,960,445 (= n*(n+1)/2, n=16790)

Diagonal interpretation:

C[i, i] = number of chunks containing concept_id = i.

Cooccurrence index mapping

For 0 <= i <= j < n:

# n = number of concepts
idx = i * n - (i * (i - 1)) // 2 + (j - i)

Read value with:

value = packed[idx]  # equals C[i, j]

Symmetry:

C[j, i] = C[i, j].

Basic dataset stats (v0)

documents (distinct doc_int_id): 14,868,862
chunks (fineweb-atlas-annotated): 95,486,049
total tokens (sum(chunk_token_count)): 10,183,028,973
reverse-index rows: 1,406,432,869
concepts: 16,790
cooccurrence max count: 79,810,673
average labels per chunk: 14.73
- content: 5.68
- tone: 6.90
- document: 1.52
- entity: 0.63

Notes and caveats

Labels are machine-generated and include noise, especially in rarer concepts.
Some chunk text is noisy/garbled source text; concept quality depends on chunk quality.
This release is intended for analysis/research and may be revised in future versions.

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; Accessed: 2026-03-04}
}