C10X/finepdfs-edu-hq

FinePDFs-Edu (English) — Filtered High-Signal Subset

This dataset is a filtered, English-only subset of HuggingFaceFW/finepdfs-edu, created to retain high-signal educational passages while reducing common PDF-extraction noise (covers/TOCs, fragmented headers/footers, OCR artifacts, mixed-language pages, and very short low-context snippets).

It is intended for training and research workflows that benefit from longer, coherent educational text extracted from PDFs.

---

At a glance

Metric	Value
Source Dataset	HuggingFaceFW/finepdfs-edu
Language	English (eng_Latn)
Total Processed	23,023,372 rows
Filtered (Kept)	5,512,514 rows (23.94%)
Rejected	17,510,858 rows (76.06%)
Total Tokens	28,252,158,603
Average Tokens	5,125
Output Shards	55
Output Size	38.37 GB

---

Why this dataset exists

FinePDFs-Edu is large and valuable, but PDF-to-text extraction naturally includes many borderline samples. In practice, many projects want a more “ready-to-train” slice with:

• Language consistency (English-only with high confidence)
• Sufficient length (avoid short fragments that behave like noise)
• Minimum educational usefulness (filter out low-signal content)

This release provides a simple, transparent filtering recipe plus reproducible logs and summary statistics.

---

Filtering strategy

Four filters were applied. Each targets a common failure mode in PDF-derived text.

1) English-only documents

• full_doc_lid == "eng_Latn"

This enforces a single dominant language at the document level.

2) High-confidence language detection

• page_average_lid_score >= 0.9

Using a page-averaged confidence score helps reduce multilingual leakage and typical OCR/segmentation issues.

Note: full_doc_lid_score is preserved in the schema for analysis, but this release uses page-level confidence as the primary LID quality signal.

3) Minimum content length

• token_count >= 512

In PDF corpora, very short samples are often headers/menus/fragmented text. A 512-token minimum strongly reduces these low-context snippets.

4) Minimum educational quality

• fw_edu_scores >= 2.0

This threshold aims to keep content that is at least moderately educational, while still preserving breadth.

---

Data quality snapshots

These binned summaries make it easy to sanity-check the retained distribution.

Language ID confidence (page_average_lid_score)

Range	Count	Percentage	Visual
0.90-0.92	246,335	4.5%	██
0.92-0.94	312,234	5.7%	██
0.94-0.96	404,114	7.3%	███
0.96-0.98	581,289	10.5%	█████
0.98-1.00	3,968,537	72.0%	███████████████████████████████████

Interpretation: The distribution is heavily concentrated in 0.98–1.00, indicating very high English confidence after filtering.

---

Token distribution

Metric	Value
Total Tokens	28,252,158,603
Average	5,125
Minimum	512
Maximum	12,113,077

Interpretation: Average length suggests medium-to-long passages (often multi-page sections). Some documents are extremely long; many pipelines will prefer to chunk or span-sample at load time.

---

Educational score notes (fw_edu_scores)

fw_edu_scores is a heuristic educational-value score. In practice, lower values often correlate with noisier extraction or weaker instructional structure, while higher values often correlate with more structured, tutorial- or textbook-like writing.

Important nuance: the rubric is oriented toward school-level educational value, so highly technical/advanced material may not always receive the highest score even if it is high quality.

---

Intended use cases

• Continued pretraining / domain adaptation toward educational writing styles
• Fine-tuning on longer instructional passages (summarization, explanation, QA, classification)
• Research on PDF extraction, educational text quality, and filtering strategies

---

Reproducibility (kept logs)

Filtering configuration (as used to produce this release):

• Dataset: HuggingFaceFW/finepdfs-edu
• Subset: eng_Latn
• Filters:
  • full_doc_lid = eng_Latn
  • page_average_lid_score >= 0.9
  • token_count >= 512
  • fw_edu_scores >= 2.0
• Output:
  • Sharded Parquet (~100,000 rows per shard)
  • Compression: zstd

---

Performance

This run was executed on a CPU-only setup:

• CPU: 4 cores / 8 threads

Stage	Duration
Dataset Load	2.29s
Processing	21,558s (~6 hours)
Total	21,560s
Processing Speed	~1,068 rows/second

---

Schema (key columns)

The dataset follows the original FinePDFs-Edu schema. The most commonly used columns are:

Column	Type	Description
text	string	Document text
id	string	Unique identifier
token_count	int64	Token count
page_average_lid	string	Page-averaged language ID
page_average_lid_score	float64	Page-averaged LID confidence
full_doc_lid	string	Document-level language ID
full_doc_lid_score	float64	Document-level LID confidence
fw_edu_scores	float64	Educational score
minhash_cluster_size	int64	Cluster size proxy (dedupe signal)
duplicate_count	int64	Duplicate count

For the complete schema, see the dataset viewer / parquet schema.

---

Storage layout

optimized/
├── README.md
├── dataset_info.json
├── shard-00000.parquet
├── shard-00001.parquet
├── ...
└── shard-00054.parquet

---

Usage examples

Fast analytics with DuckDB

import duckdb

con = duckdb.connect()

stats = con.execute("""
    SELECT
        COUNT(*) AS rows,
        AVG(token_count) AS avg_tokens,
        AVG(page_average_lid_score) AS avg_lid,
        AVG(fw_edu_scores) AS avg_edu
    FROM 'optimized/*.parquet'
""").fetchall()

print(stats)

# Optional: create a cleaner slice
con.execute("""
    COPY (
        SELECT id, text
        FROM 'optimized/*.parquet'
        WHERE fw_edu_scores >= 3.0
        AND token_count >= 1024
    )
    TO 'train_slice.parquet' (FORMAT PARQUET);
""")

---

Citation

If you use this dataset, please cite the original FinePDFs dataset:

@misc{kydlicek2025finepdfs,
      title={FinePDFs},
      author={Kydlíček, Hynek and Penedo, Guilherme and von Werra, Leandro},
      year={2025},
      publisher = {Hugging Face},
      howpublished = {\url{https://huggingface.co/datasets/HuggingFaceFW/finepdfs-edu}}
}

---

License

This filtered dataset inherits the ODC-By v1.0 license from the original FinePDFs-Edu dataset.

---

Changelog

• 2026-01-18: Initial filtered release (English-only, high-confidence LID, min length 512 tokens, min edu score 2.0)

---

Generated by: FinePDFs-Edu Dataset Streamer
Generated at: 2026-01-18