StentorLabs/StenCore-PDF

StenCore — FinePDFs-Edu Curated

By StentorLabs

StenCore is StentorLabs' first dataset release — a quality-first resource built from the ground up for training language models. Derived from HuggingFaceFW/finepdfs-edu, every document passed a 14-stage automated curation pipeline including heuristic filtering, language ID, PII redaction, toxicity screening, eval decontamination, multi-strategy deduplication, KenLM + neural perplexity scoring, and domain reweighting.

⚠️ Privacy & Copyright Notice: This dataset was produced by automated pipelines that include regex-based PII redaction and heuristic filters. Automated redaction may miss personal data (false negatives) and may over-redact (false positives). No claim of complete PII removal is made. Do not use this dataset in systems that surface individual personal data. Review and clear upstream site-level licenses before redistributing or deploying models trained on this data. For takedown or privacy requests, open an issue or contact stentorlabs@gmail.com with the affected doc identifier and evidence.

---

Quick Stats

Property	Value
Pipeline	StenCore v2026.03
Source	HuggingFaceFW/finepdfs-edu (3 Parquet files, ~8.9 GB)
Docs in → out	584,000 → 149,000 (~37.78% keep rate)
Size (uncompressed)	447,088,827 bytes (~447.09 MB)
Avg doc size	2,988 bytes (750 words)
Synthetic docs	0 (100% human-authored)
Language	English (en)
Platform	Kaggle CPU
Total runtime	~9.47 hours wall / ~12+ hours CPU

---

Usage

from datasets import load_dataset

ds = load_dataset("StentorLabs/stencore")

# Filter to highest-quality documents
top = ds["train"].filter(lambda r: r["cqf_score"] >= 0.5)
print(len(top))
print(top[0]["text"][:500])

---

Intended & Disallowed Uses

✅ Intended uses

• LLM pre-training and fine-tuning on English educational text
• Quality filtering and data curation research
• Domain reweighting and curriculum learning experiments

🚫 Disallowed uses

• Systems that surface or process personal information
• Redistribution without reviewing upstream licensing (Apache 2.0 requires attribution; check site-level licenses for individual documents)
• Any production deployment without independent legal and privacy review

---

Pipeline Summary

StenCore (curate_2026 mode) is a fully automated curation pipeline (v2026.03) running 14 sequential stages from raw Parquet to a publish-ready HuggingFace dataset.

Every document in this dataset passed all of the following:

1. Columnar pre-filter — DuckDB SQL-level filter over raw Parquet
2. Adaptive quality thresholds — learned from the data, not hand-tuned (per-register: formal / conversational / technical)
3. Language ID gate — fastText or stopword/script heuristic (en only)
4. Heuristic quality filter — word count, alpha ratio, stopword hits, punctuation ratio, repetition ratios, avg word length
5. PII redaction — regex detection of emails, phones, SSNs, IPs, card numbers, IBANs, API keys
6. Toxicity screening — lexical axis scoring, drop policy
7. Eval decontamination — exact, 8-gram overlap, char 3-gram, SimHash checks against benchmark sets
8. Exact deduplication — SHA-1 fingerprint (in-memory)
9. MinHash near-dedup — token shingle LSH (autotuned to 25% sample rate)
10. CQF quality gate — top 85% by quality score (threshold: CQF ≥ 0.4167)
11. KenLM + neural perplexity gate — edugp/kenlm Wikipedia model + SmolLM2-135M (excess mode, ref mean: 916.24)
12. Cluster rehydration — restores high-quality representatives from dedup-dropped clusters
13. Domain reweighting — per-domain weights learned from proxy model evaluation
14. Mix optimization + final write — domain-quota sampling, curriculum ordering, shard + merkle output

📋 Full Stage-by-Stage Breakdown

Pre-Stage: Environment & Model Setup

KenLM Scorer

• Repo: edugp/kenlm | Corpus: wikipedia/en
• Snapshot: 3fbe35c83b1a39f420a345b7c96a186c8030d834
• Mode: first_pass — KenLM prefilters, SmolLM2 rescores the subset

Neural Reference LM

• Model: HuggingFaceTB/SmolLM2-135M (torchao int8_weight_only)
• torch.compile: Disabled (CPU stability)

Effective runtime profile:

light_mode=False  input_cap=400  bootstrap_docs=400
stage_timeout_s=90  stream_chunk=32768  ppl_batch=192
ppl_workers=1  prefix_tokens=256  kenlm_mode=first_pass

12h target profile:

hf_input_target_gb=7.50  hf_input_max_files=64
cqf_keep_web=0.85  prior_keep=0.90  web_ratio=0.95
strict_ratio=False  dedup=memory  autotune_grace=10000

---

Stage 1: resolve_web_inputs

Wall: 53.47s | CPU: 143.36s | Ratio: 2.68×

DuckDB columnar prefilter over 3 Parquet files → stage_columnar_prefilter.parquet. High CPU/wall ratio confirms parallel DuckDB execution.

---

Stage 2: adapt_multilingual_profiles

Wall: 25.48s | CPU: 25.39s | Ratio: 1.00×

Draws a pilot sample and learns per-register quality threshold parameters from the data. Three registers profiled: en:formal (1,854 samples), en:conversational (2,993 samples), en:technical. Single-threaded statistical pass.

---

Stage 3: optimize_threshold_profiles

Wall: 86.79s | CPU: 138.40s | Ratio: 1.59×

Grid search over code/math profile scales, ranked by proxy quality metric. Selects best-performing threshold configuration.

---

Stage 4: cqf_seed_verification_loop

Wall: 0.001s | CPU: 0.001s

Retrains CQF fastText at multiple thresholds, picks best by proxy quality. Completed near-instantly — seed was valid, no remediation needed.

---

Stage 5: stage_web_candidate_pass

Wall: 27,162.57s (7.5h) | CPU: 42,229.32s | Ratio: 1.55×

Main filtering stage. Sequential gates per document:

1. Quick prefilter (min words / alpha)
2. HF-specific input filters (English subset, EDU v2, strip code fences)
3. Source policy (host allow/deny, ToS risk, license allowlist)
4. Line cleaner (boilerplate, nav, HTML artifacts, duplicate lines)
5. Domain routing (code / math / prose)
6. Language ID gate (+ English noise guard, code bypass)
7. Domain-aware heuristic filter (adaptive thresholds from Stage 2/3)
8. PII redaction
9. Toxicity screening
10. Eval decontamination (exact, 8-gram, char 3-gram, SimHash)
11. CQF scoring
12. Exact dedup (SHA-1, in-memory)
13. MinHash near-dedup (autotuned)
14. Semantic dedup (autotuned to none)

Kept: 220,635 / 584,000 (37.78%). Dropped records to semantic dup pool for potential rehydration.

---

Stage 6 (observed): mid_proxy_stage3

Wall: 115.50s | CPU: 160.80s | Ratio: 1.39×

Intermediate proxy scoring pass over quality-filtered candidates.

---

Stage 7: stage_web_quality_and_perplexity

Wall: 300.05s | CPU: 282.49s | Ratio: 0.94×

1. CQF threshold gate — keeps top 85% (CQF ≥ 0.4167)
2. Multi-objective property minima — per-property floor checks
3. Prior noise gate — filters statistical outliers vs. quality prior
4. Hybrid disagreement trigger — routes CQF/secondary disagreements to full neural perplexity
5. KenLM + SmolLM2 perplexity gate — excess mode, ref mean 916.24

22,450 documents scored. Avg perplexity: 902.02 (stored as excess perplexity — relative to the 916.24 calibrated mean, not absolute).

---

Stage 8: rehydrate_clusters

Wall: 40.42s | CPU: 40.19s | Ratio: 0.99×

Reads the semantic dup pool. Re-adds top-quality cluster representatives (ranked by FineWeb2-like weighted formula). Optional MMR diversity selection. Single-threaded.

---

Stage 9 (observed): mid_proxy_stage5

Wall: 102.92s | CPU: 143.10s | Ratio: 1.39×

Second intermediate proxy scoring pass after cluster rehydration.

---

Proxy Eval & Domain Reweighting (~60 min observed gap)

Per-domain proxy scores aggregated → per-domain reweighting coefficients learned. Hundreds of source domains reweighted. Examples:

Domain	Weight
apps.oregonlegislature.gov	1.0811
fcaresources.com	1.0159
ymcawnc.org	0.9999
www2.cs.arizona.edu	0.9585
echalk-slate-prod.s3.amazonaws.com	0.8882

---

Stage 9: build_synthetic_pool

Wall: 0.003s — No teacher LLM configured. Pool: 0 documents.

Stage 10: stage_synthetic_filter

Wall: 0.002s — No-op (empty pool).

---

Stage 11: mix_optimization

Wall: 175.34s | CPU: 214.87s | Ratio: 1.23×

Proxy-evaluated search over web/synth mixing ratios. Applies domain weights from proxy eval. Produces final document selection plan.

---

Stage 12: final_mix_and_write

Wall: 2,311s (38.5 min)

Domain quota enforcement → domain-weighted acceptance sampling → final exact dedup → optional curriculum ordering (easy→hard by CQF/perplexity) → streaming JSONL + Parquet write → SHA-256 + merkle root → optional deterministic sharding.

Docs Written	Bytes	Avg Bytes/Doc
1,000	2,783,132	2,783
149,000	447,088,827	2,999

Stages 13–14: proxy_eval + hf_push

Final proxy metric gate enforcement → auto-upload to HuggingFace Hub.

---

Adaptive Quality Thresholds

📐 Full learned threshold values (en:formal and en:conversational)

These parameters were learned from the data in Stage 2 and optimized in Stage 3. All values are exact as logged.

en:formal — 1,854 bootstrap samples

Parameter	Value
min_words	30
max_words	2,000
min_stopwords	2
max_line_punct_ratio	0.1111111111111111
max_word_repeat_3gram_ratio	0.30466436237947997
max_char_repeat_5gram_ratio	0.35
min_alpha_ratio	0.5967419247419248
min_avg_word_len	4.157040378006873
max_avg_word_len	6.216977322149734

en:conversational — 2,993 bootstrap samples

Parameter	Value
min_words	18
max_words	2,000
min_stopwords	2
max_line_punct_ratio	0.10838961038961101
max_word_repeat_3gram_ratio	0.19476069102237326
max_char_repeat_5gram_ratio	0.35
min_alpha_ratio	0.6542321503584156
min_avg_word_len	4.098954647914038
max_avg_word_len	6.0

en:technical

Separate profile applied; full values truncated in logs. Expected to have higher tolerance for non-alphabetic characters (equations, code, symbols) and relaxed stopword requirements.

---

Deduplication Autotune

⚙️ Runtime autotune event log

The autotune system fires after 10,000 documents (grace period). All 6 events occurred in a 5,000-document window — the system converges aggressively.

Event	Docs Seen	Throughput	Action
Start	0	—	sem=hybrid, embed=0.0200, minhash=1.000
#1	10,000	21.81 docs/s	embed_sample → 0.0050
#2	11,000	21.92 docs/s	embed_sample → 0.0020
#3	12,000	22.04 docs/s	semantic_mode → minhash
#4	13,000	22.13 docs/s	minhash_sample → 0.500
#5	14,000	22.31 docs/s	minhash_sample → 0.250
#6	15,000	22.57 docs/s	semantic_mode → none
Stable	584,000	21.5 docs/s	Final: sem=none, embed=0.0020, minhash=0.250

Throughput gain: ~1.8× (12.3 → 22 docs/s). MinHash at 25% sampling will miss some near-duplicate pairs — a deliberate throughput/precision tradeoff accepted by the autotune system based on observed duplicate density.

---

Perplexity Scoring

📊 KenLM + SmolLM2 scoring details

Configuration

• KenLM model: edugp/kenlm, wikipedia/en, snapshot 3fbe35c83b1a39f420a345b7c96a186c8030d834
• Neural LM: HuggingFaceTB/SmolLM2-135M (torchao int8_weight_only)
• Mode: first_pass — KenLM prefilters all docs; SmolLM2 rescores near-boundary subset
• Scoring mode: excess — score = raw_perplexity - reference_mean
• Reference mean: 916.2404 (calibrated fresh on 256 bootstrap docs, fit_new)

Scoring progress

Docs Scored	Last Doc PPL	Running Avg PPL
50	7,442.60	218.38
22,450	8,182.66	902.02

High individual perplexity values (>7,000) are expected for math-heavy or notation-rich educational text under a Wikipedia-trained model. The excess mode partially normalizes this. Running mean converges to ~902 across the full scored set.

---

Runtime & Timing

⏱️ Full per-stage timing table

Stage	Wall Time	CPU Time	CPU/Wall
Model/env setup	~167s	—	—
resolve_web_inputs	53.47s	143.36s	2.68×
adapt_multilingual_profiles	25.48s	25.39s	1.00×
optimize_threshold_profiles	86.79s	138.40s	1.59×
cqf_seed_verification_loop	0.001s	0.001s	1.00×
stage_web_candidate_pass	27,162.57s	42,229.32s	1.55×
stage_web_quality_and_perplexity	300.05s	282.49s	0.94×
mid_proxy_stage3 (observed)	115.50s	160.80s	1.39×
rehydrate_clusters	40.42s	40.19s	0.99×
mid_proxy_stage5 (observed)	102.92s	143.10s	1.39×
Proxy / domain reweighting	~3,590s	—	—
build_synthetic_pool	0.003s	0.005s	—
stage_synthetic_filter	0.002s	0.002s	—
mix_optimization	175.34s	214.87s	1.23×
final_mix_and_write	~2,311s	—	—
TOTAL	~34,088s	~43,377s+	~1.27×

Candidate pass throughput: 12.3 docs/s (initial) → ~22 docs/s (post-autotune, **1.8× gain**).

---

Limitations

• PDF extraction artifacts — OCR artifacts, broken equations, and malformed tables may be present despite filtering.
• Residual PII — Automated regex redaction does not guarantee complete PII removal. Do not use for systems that surface personal information.
• Copyright — Source PDFs may carry individual site-level licenses. Apache 2.0 requires attribution; verify upstream licensing for your use case.
• KenLM Wikipedia bias — Math-heavy or highly technical documents may be underrepresented due to high perplexity under a Wikipedia-trained model.
• ~62% rejection rate — Some valid educational content may have been dropped due to heuristic threshold mismatch (e.g., table-heavy or equation-dense formatting).
• English only — Pipeline profiled en:formal, en:conversational, and en:technical registers only.
• No synthetic data — This run did not use the synthetic generation system. Dataset is 100% source text.
• MinHash 25% sampling — Post-autotune dedup will miss some near-duplicate pairs.

---

Licensing & Citation

Released under Apache 2.0. Attribution required. Derived from HuggingFaceFW/finepdfs-edu — review upstream licensing before use.

@dataset{stencore_finepdfs_edu_curated,
  title        = {StenCore: FinePDFs-Edu Curated},
  author       = {StentorLabs},
  year         = {2026},
  note         = {StentorLabs' first dataset. StenCore pipeline v2026.03.
                  584k docs in, 149k out. Adaptive heuristics, PII redaction,
                  toxicity/decontam gates, MinHash + KenLM + neural perplexity,
                  CQF scoring, proxy domain reweighting.},
  howpublished = {\url{https://huggingface.co/datasets/StentorLabs/stencore}}
}

@dataset{fineweb_finepdfs_edu,
  author       = {HuggingFace FineWeb Team},
  title        = {FinePDFs-Edu},
  howpublished = {\url{https://huggingface.co/datasets/HuggingFaceFW/finepdfs-edu}}
}

Contact: StentorLabs@gmail.com — for takedown requests, privacy concerns, or feedback.

---

Made with ❤️ by StentorLabs