NoeFlandre/fineweb-polygons

FineWeb Polygons

GitHub repository · Hugging Face dataset

The public dataset catalog maps every immutable V1–V9 file to its country split, standalone version README, and reproducibility manifest. The readable GitHub catalog and HF metadata README link back to both projects.

FineWeb Polygons finds high-confidence FineWeb documents that are directly tied to OpenStreetMap polygons. V1 through V4 use Monaco; V5 through V9 run the experiment for Monaco and Liechtenstein. Raw OSM extracts are kept outside the repository at:

/Volumes/Seagate M3/projects/fineweb-polygons/raw/monaco-latest.osm.pbf and /Volumes/Seagate M3/projects/fineweb-polygons/raw/liechtenstein-latest.osm.pbf.

The version ID is part of the public contract. Existing version IDs are not silently changed; a behavior change gets a new version ID.

Version contracts

V1 — named polygon exact matching

Run with --retrieval-version v1. From the PBF, V1 keeps every named closed way and polygon relation. A FineWeb document is kept when a polygon name and Monaco or Principality of Monaco occur in the text or URL. Matching is case-insensitive and exact after normalization. The output keeps the complete FineWeb text.

The uploaded V1 file is the original excerpt-only publication and therefore has text_excerpt and url_excerpt rather than a text column. New V1 runs made by the current code keep the complete text; the old public file is preserved at its original path and is not silently replaced.

V2 — meaningful in-boundary exact matching

Run with --retrieval-version v2. From the raw PBF, V2 keeps named area objects inside the Monaco admin_level=8 city boundary. It rejects names shorter than three characters, numeric-only names, and labels without letters, then deduplicates normalized names.

For FineWeb, a URL name match is enough. A text name match is kept only when the same text also contains Monaco or Principality of Monaco. The output keeps the complete FineWeb text, URL, and evidence fields.

V3 — all meaningful polygon areas with strict URL-and-text matching

Run with --retrieval-version v3. From the raw PBF, V3 keeps every valid area produced from a closed way or relation. It uses only the main name tag, rejects names shorter than three characters, numeric-only names, and labels without letters, then deduplicates normalized names.

A FineWeb document is kept only when the polygon name appears in the URL and the polygon name plus Monaco or Principality of Monaco appear in the text. Final evidence is deduplicated per polygon and document, using the FineWeb ID when available and otherwise the URL plus a hash of the complete text. The output keeps the complete FineWeb text, URL, and evidence fields.

V4 — all meaningful polygon areas with text-only matching

Run with --retrieval-version v4. V4 uses the same polygon profile as V3: every valid named area from a closed way or polygon relation, with the same name cleanup and normalized-name deduplication.

A FineWeb document is kept when the polygon name and Monaco or Principality of Monaco both appear in the text. The URL is retained as metadata and evidence, but it is not a selection condition. Final evidence is deduplicated per polygon and document, and the output keeps the complete FineWeb text.

The exact definitions are stored in src/fineweb_polygons/versions.py. Every run manifest copies the selected definition and hashes it as part of the configuration, so a changed definition cannot silently resume an old run. See the version guide for the same contract in a readable format.

V5 - specific polygon areas with country-in-text matching

Run with the v5 retrieval version and pass the relevant country name. V5 starts with every meaningful named area from the PBF, just like V3 and V4. It does not search for or require a country boundary.

Before retrieving documents, V5 counts each normalized polygon name:

1. A name must occur in only one OSM area.
2. It must occur in no more than 0.1% of the FineWeb shard's documents.
3. The configured country name is context only; it is never kept as a polygon candidate.

A document is kept only when the selected polygon name and the exact country name both occur in the FineWeb text. The URL is retained as evidence but does not select a document. Matching is case-insensitive and exact after normalization. Results keep the complete FineWeb text and are deduplicated per polygon and FineWeb document.

The first V5 pass writes a name-frequency.json file inside the run directory. Its input fingerprints, counts, cutoff, and retained names are recorded there and its SHA-256 is copied into the run manifest. A restart reuses this artifact and skips the frequency pass.

Foundation contract

• uv owns the locked Python environment.
• Ruff, ty, pytest, mutation testing, and a CRAP gate are wired into local and CI commands.
• Docker and MkDocs Material are configured from the start.
• LICENSE and CITATION.cff are public project artifacts.
• Raw input, run manifests, checkpoints, logs, and generated artifacts stay on the Seagate project volume.
• V1/V2/V3/V4/V5/V6 processing is resumable in chunks of 32 Parquet row groups, appends structured JSON logs, and records input/configuration fingerprints in a manifest. V5 and V6 also checkpoint their name-frequency pass. V7, V8, and V9 are content-fingerprinted, atomic post-processing runs that can reuse completed outputs.
• Every result is tied to an explicit retrieval version; new retrieval behavior must use a new version ID.
• The implementation uses small, deep modules with stable interfaces and YAGNI scope.

V1/V2/V3/V4/V5/V6 intentionally skip aliases, fuzzy matching, embeddings, and classifiers. V7 only segments already selected V6 documents; it does not change document selection. V8 filters already selected V7 documents with a fixed topic vocabulary at document level. V9 filters V8 sentences with the same vocabulary near polygon-name evidence; it does not add new polygon names or run an LLM. They produce evidence JSONL on the Seagate; the raw shard and model cache are never uploaded.

Public tiny-shard artifacts

The public dataset contains the filtered evidence from the first shard:

V5 and V6 publish one split for Monaco and one split for Liechtenstein. Both use the same full-text evidence schema and can be inspected independently in the Hugging Face viewer.

• V1: data/v1/monaco-v1-10bt-000-v3.jsonl
• V2: data/v2/monaco-v2-10bt-000-v2.jsonl
• V3: data/v3/monaco-v3-10bt-000-v1.jsonl
• V4: data/v4/monaco-v4-10bt-000-v1.jsonl
• V5 Monaco: data/v5/monaco-v5-10bt-000-v3.jsonl
• V5 Liechtenstein: data/v5/liechtenstein-v5-10bt-000-v2.jsonl
• V6 Monaco: data/v6/monaco-v6-10bt-000-v1.jsonl
• V6 Liechtenstein: data/v6/liechtenstein-v6-10bt-000-v1.jsonl
• V7 Monaco: data/v7/monaco-v7-10bt-000-v1.jsonl
• V7 Liechtenstein: data/v7/liechtenstein-v7-10bt-000-v1.jsonl
• V8 Monaco: data/v8/monaco-v8-10bt-000-v1-topic.jsonl
• V8 Liechtenstein: data/v8/liechtenstein-v8-10bt-000-v1-topic.jsonl
• V9 Monaco: data/v9/monaco-v9-10bt-000-v1-topic-sentences.jsonl
• V9 Liechtenstein: data/v9/liechtenstein-v9-10bt-000-v1-topic-sentences.jsonl

These are evidence records, not a copy of the raw FineWeb shard. The published schemas are intentionally documented separately:

• V1 is the original excerpt-only release: it contains text_excerpt and url_excerpt, but no text field.
• V2 is the full-text release: it contains the complete FineWeb document in text, plus the short preview fields.
• V3 is the strict URL-and-text release: it contains the complete FineWeb document in text, plus the short preview fields and deduplicated evidence.
• V4 is the text-only release: it contains the complete FineWeb document in text, plus the short preview fields and deduplicated evidence. Its URL is retained for inspection but does not decide selection.

The V3 first-shard run read 1,048,581 FineWeb documents from a 2.0 GB Parquet file. The shard contains 539,338,878 whitespace-separated words (and a FineWeb token_count sum of 726,306,534). V3 retained 77 final evidence records across 7 polygon names.

The V4 first-shard run scanned the same 1,048,581 documents and retained 1,282 deduplicated evidence records across 42 polygon names. It contains 1,205 records selected from text alone and 77 where the URL also contains the name; all 1,282 records retain the complete FineWeb text. Because V4 intentionally removes the URL condition, the result is a higher-recall baseline and is still broad for generic names: 1,067 records use the polygon name Monaco. This is documented as an experiment, not as a claim that every text-only match is already semantically correct.

The corrected V5 runs use the same 1,048,581-document shard. Monaco started with 741 normalized names, kept 713 after removing 23 repeated OSM names, four names above the 0.1% FineWeb cutoff, and the country name itself, then wrote 61 deduplicated records across 33 polygon names. Liechtenstein started with 565 names, kept 524, and wrote 12 deduplicated records across 11 polygon names. Every result has the complete text; 57 Monaco and 10 Liechtenstein records matched only in text, while the remaining four and two also had the name in the URL. The country name appeared only as context.

These are still lexical results, not semantic proof. Manual review found some false positives such as Buckingham, Bel Air, P 1, and Maxi. Frequency filtering removes very common labels but cannot know whether a short name is the intended place; a later version will need stronger evidence for that.

V6 — local text-span evidence

V6 keeps the V5 polygon and frequency rules, then requires the selected polygon name and configured country name to be within 500 normalized characters in the FineWeb text. The URL is not a selection condition. Each V6 row keeps the full FineWeb text, the closest normalized distance, and the original-text sentence for the polygon name and country name. V6 does not include text_excerpt or url_excerpt columns. The published Monaco and Liechtenstein files are separate viewer splits under data/v6/.

V7 — exact sentence lists from V6

V7 is a post-processing version, not a new retrieval rule. It reads the published V6 rows, keeps every V6 field including the complete text, and adds an ordered sentences list produced by sat-3l-sm from wtpsplit. The splitter uses split_on_input_newlines=false and strip_whitespace=false; the pipeline requires ''.join(sentences) == text, so it cannot silently rewrite the source document. Monaco has 45 rows and 4,569 sentences; Liechtenstein has 6 rows and 328 sentences. The model and segmentation settings are recorded in each V7 manifest. V7 does not upload the model or any raw FineWeb/OSM data.

The code and manifests still preserve the retrieval definition for each version; new output should use a new artifact path rather than overwriting a published file.

V8 — topic-vocabulary filtering from V7

V8 is a post-processing version, not a new polygon retrieval rule. It reads each V7 JSONL artifact and keeps a complete V7 row only when its full text contains at least one of the 136 approved strong topic terms. Matching is case-insensitive, uses Unicode NFKC normalization and whole-word boundaries, and never examines the URL. The same vocabulary is used for Monaco and Liechtenstein; it is not country-aware.

V8 preserves every kept V7 field, including the complete text and ordered sentences list. It records the source, result, and vocabulary SHA-256 values, term categories, row counts, and matching settings in its manifest. The vocabulary JSON and both country manifests are published as HF metadata; raw FineWeb, OSM, and model files remain on the Seagate.

On the first shard, V8 kept 39 of 45 Monaco V7 documents and filtered 6; the kept rows cover 25 polygon names and contain 4,472 sentences. It kept all 6 Liechtenstein V7 documents, covering 5 polygon names and 328 sentences. Category counts overlap because one document can match several categories.

V9 — sentence-level topic filtering from V8

V9 is a post-processing version. It reads V8 rows and examines each original sentence using the same 136-term vocabulary. A sentence is kept only when it contains at least one whole-word topic term and is at most two sentence positions from a sentence containing the polygon name. This bounded window keeps place context without treating an unrelated topic word elsewhere in a long document as evidence. The country remains an audit anchor; V8 has already required the polygon name and country to be within 500 normalized characters.

V9 preserves the complete text and original sentences list, then adds relevant_sentences as a plain list of sentence strings for quick review. The aligned relevant_sentence_metadata list stores the original sentence index, matched terms, matched categories, topic counts, and polygon/country sentence distances without repeating the sentence text. Rows with no local topic sentence are filtered. The output and manifest are atomic and reusable when the V8 input, vocabulary, settings, and result hash match. This is V9 output schema version 2; the selection rule is unchanged.

On the first shard, V9 kept 29 of 39 Monaco V8 rows and wrote 93 relevant sentences. It kept 4 of 6 Liechtenstein V8 rows and wrote 9 relevant sentences. The polygon-name evidence was in the same sentence for 33 Monaco sentences and 3 Liechtenstein sentences; the remaining matches were one or two sentence positions away.

First-shard run

Keep the Hugging Face cache, virtual environment, raw data, and run outputs on the Seagate:

export HF_HUB_CACHE="/Volumes/Seagate M3/projects/fineweb-polygons/cache/huggingface-v7/hub"
export UV_CACHE_DIR="/Volumes/Seagate M3/projects/fineweb-polygons/cache/uv-cleanup"
export UV_PROJECT_ENVIRONMENT="/Volumes/Seagate M3/projects/fineweb-polygons/.venvs/fineweb-polygons-v8"

uv sync --locked
hf download HuggingFaceFW/fineweb \
  --repo-type dataset \
  --include "sample/10BT/000_00000.parquet" \
  --local-dir "/Volumes/Seagate M3/projects/fineweb-polygons/raw/fineweb"

uv run fineweb-polygons scan \
  --pbf "/Volumes/Seagate M3/projects/fineweb-polygons/raw/monaco-latest.osm.pbf" \
  --shard "/Volumes/Seagate M3/projects/fineweb-polygons/raw/fineweb/sample/10BT/000_00000.parquet" \
  --run-id v1-10bt-000-v3

uv run fineweb-polygons scan \
  --pbf "/Volumes/Seagate M3/projects/fineweb-polygons/raw/monaco-latest.osm.pbf" \
  --shard "/Volumes/Seagate M3/projects/fineweb-polygons/raw/fineweb/sample/10BT/000_00000.parquet" \
  --run-id v2-10bt-000-v2 \
  --retrieval-version v2

uv run fineweb-polygons scan \
  --pbf "/Volumes/Seagate M3/projects/fineweb-polygons/raw/monaco-latest.osm.pbf" \
  --shard "/Volumes/Seagate M3/projects/fineweb-polygons/raw/fineweb/sample/10BT/000_00000.parquet" \
  --run-id v3-10bt-000-v1 \
  --retrieval-version v3

uv run fineweb-polygons scan \
  --pbf "/Volumes/Seagate M3/projects/fineweb-polygons/raw/monaco-latest.osm.pbf" \
  --shard "/Volumes/Seagate M3/projects/fineweb-polygons/raw/fineweb/sample/10BT/000_00000.parquet" \
  --run-id v4-10bt-000-v1 \
  --retrieval-version v4

Each run stores chunk checkpoints and a manifest under its run ID, logs under logs/, and merged evidence under artifacts/, all below the Seagate project root. One chunk opens the Parquet shard once and covers at most 32 row groups.

V5 commands for the two corrected runs:

uv run fineweb-polygons scan \
  --pbf "/Volumes/Seagate M3/projects/fineweb-polygons/raw/monaco-latest.osm.pbf" \
  --shard "/Volumes/Seagate M3/projects/fineweb-polygons/raw/fineweb/sample/10BT/000_00000.parquet" \
  --run-id v5-monaco-10bt-000-v3 \
  --retrieval-version v5 \
  --country-name "Monaco"

uv run fineweb-polygons scan \
  --pbf "/Volumes/Seagate M3/projects/fineweb-polygons/raw/liechtenstein-latest.osm.pbf" \
  --shard "/Volumes/Seagate M3/projects/fineweb-polygons/raw/fineweb/sample/10BT/000_00000.parquet" \
  --run-id v5-liechtenstein-10bt-000-v2 \
  --retrieval-version v5 \
  --country-name "Liechtenstein"

V7 sentence segmentation reads those V6 artifacts and writes new artifacts; all paths below remain on the Seagate:

export HF_HUB_CACHE="/Volumes/Seagate M3/projects/fineweb-polygons/cache/huggingface-v7/hub"

uv run fineweb-polygons segment-v7 \
  --data-root "/Volumes/Seagate M3/projects/fineweb-polygons" \
  --input "/Volumes/Seagate M3/projects/fineweb-polygons/artifacts/v6-monaco-10bt-000-v1-matches.jsonl" \
  --output "/Volumes/Seagate M3/projects/fineweb-polygons/artifacts/v7-monaco-10bt-000-v1-sentences.jsonl" \
  --manifest "/Volumes/Seagate M3/projects/fineweb-polygons/runs/v7-monaco-10bt-000-v1/manifest.json"

uv run fineweb-polygons segment-v7 \
  --data-root "/Volumes/Seagate M3/projects/fineweb-polygons" \
  --input "/Volumes/Seagate M3/projects/fineweb-polygons/artifacts/v6-liechtenstein-10bt-000-v1-matches.jsonl" \
  --output "/Volumes/Seagate M3/projects/fineweb-polygons/artifacts/v7-liechtenstein-10bt-000-v1-sentences.jsonl" \
  --manifest "/Volumes/Seagate M3/projects/fineweb-polygons/runs/v7-liechtenstein-10bt-000-v1/manifest.json"

V8 topic filtering reads those V7 artifacts and keeps only rows whose full text contains at least one approved vocabulary term:

uv run fineweb-polygons filter-v8 \
  --data-root "/Volumes/Seagate M3/projects/fineweb-polygons" \
  --input "/Volumes/Seagate M3/projects/fineweb-polygons/artifacts/v7-monaco-10bt-000-v1-sentences.jsonl" \
  --output "/Volumes/Seagate M3/projects/fineweb-polygons/artifacts/v8-monaco-10bt-000-v1-topic.jsonl" \
  --manifest "/Volumes/Seagate M3/projects/fineweb-polygons/runs/v8-monaco-10bt-000-v1/manifest.json" \
  --vocabulary "/Volumes/Seagate M3/projects/fineweb-polygons/v8-topic-vocabulary-v1.json"

uv run fineweb-polygons filter-v8 \
  --data-root "/Volumes/Seagate M3/projects/fineweb-polygons" \
  --input "/Volumes/Seagate M3/projects/fineweb-polygons/artifacts/v7-liechtenstein-10bt-000-v1-sentences.jsonl" \
  --output "/Volumes/Seagate M3/projects/fineweb-polygons/artifacts/v8-liechtenstein-10bt-000-v1-topic.jsonl" \
  --manifest "/Volumes/Seagate M3/projects/fineweb-polygons/runs/v8-liechtenstein-10bt-000-v1/manifest.json" \
  --vocabulary "/Volumes/Seagate M3/projects/fineweb-polygons/v8-topic-vocabulary-v1.json"

License and upstream data

The dataset and project artifacts use the Open Data Commons Attribution License (ODC-By) v1.0, the same license shown on the FineWeb dataset card. FineWeb also notes that its Common Crawl source is subject to Common Crawl's terms of use; downstream releases must preserve applicable upstream notices and rights.

Development

uv sync
just qa
just mutation

See the development guide and foundation architecture for the current scope.