Mati83moni/HybridMoE-Training-Dataset-v1

📊 HybridMoE Training Dataset v1

Multilingual EN-PL Pre-Training Corpus · ~9 GB · ~2.4B Tokens · Cleaned & Audited

The training data behind HybridMoE Titan v1 — curated, cleaned, and corruption-free.

🧠 Model v1 · 🚀 Model v2 · 📄 Paper · 🌐 Project Page

---

📦 Dataset Overview

Total Size	~9 GB (raw text)
Estimated Tokens	~2.4 billion
Files	51 plain text files
Languages	English (primary) + Polish
License	CC BY-SA 4.0
Format	Plain text, one document per line (double-newline separated)

---

📚 Sources

Source	Language	Size	Files	Description
🎓 FineWeb-Edu	EN	~3.5 GB	~15 files	High-quality educational web text (curated by HuggingFace)
🌐 FineWeb	EN	~2.5 GB	~15 files	General web text with quality filters
🇵🇱 Wikipedia PL	PL	~1.5 GB	~15 files	Polish Wikipedia article dumps
📖 Wolne Lektury	PL	~323 MB	1 file	7,473 Polish public-domain literary works
📝 Polski Dodatek	PL	~200 MB	1 file	Supplementary Polish text (cleaned)

---

🧹 Data Cleaning & Quality Audit

The Wolne Lektury Problem

During Titan v2 preparation, a full corruption audit revealed serious issues in the Wolne Lektury source:

Issue	Count
🖼️ JPEG book covers embedded as binary in text	306 images
❓ U+FFFD replacement characters	8.1 million
🚫 Null bytes	190,000
💥 Corrupt lines	488,000 (17.6% of file)
📚 Books affected	297 / 7,473

✅ Cleaning Result

All corruption was removed while preserving every book:

Before: 2,773,128 lines · 488,423 corrupt (17.6%)
After:  2,284,705 lines · 0 corrupt (0.0%)
Books:  7,473 / 7,473 preserved ✅

💡 The cleaned file (wolne_lektury_dataset_FINAL.txt, 323 MB) is included in this dataset.

---

🔧 Preprocessing Pipeline

Raw Sources
    │
    ├─ FineWeb-Edu/FineWeb ──→ Quality-filtered by HuggingFace
    │                          Deduplicated
    │                          Split into ~4 GB shards
    │
    ├─ Wikipedia PL ──────────→ Article extraction
    │                          Infobox/citation removal
    │                          Split into shards
    │
    └─ Wolne Lektury ────────→ Book-boundary detection
                               Binary artifact removal (306 JPEGs)
                               U+FFFD / null byte cleaning
                               Line-level validation
                               ✅ 0 corruption output

Tokenisation (Titan v1)

Tokeniser	GPT-2 BPE (50,258 tokens)
Loader	PackedDataset — zero padding waste
Sequence length	4,096 tokens
Shards	200 validation, rest for training

Tokenisation (Titan v2)

Tokeniser	Custom BPE (64,000 tokens) with native Polish support
Strategy	stream_tokenize.py with memory-efficient streaming
Domain mixing	Stratified batch sampling across sources

---

📏 Usage

This dataset is provided as raw text files. To use with HybridMoE Titan:

# Example: load and tokenise a shard
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("gpt2")

with open("fineweb_edu_shard_0000.txt", "r") as f:
    for doc in f.read().split("\n\n"):
        tokens = tokenizer.encode(doc)
        # ... your training pipeline

---

⚠️ Known Issues

• GPT-2 tokeniser fragments Polish diacritics (ą, ę, ź, ż → 2-3 tokens each) — addressed in Titan v2 with custom 64K tokeniser
• Domain imbalance — English text dominates (~70%); Polish sources are ~30%
• No deduplication across sources — potential overlap between FineWeb and Wikipedia

---

📬 Author

Mateusz Piesiak — Project Inkblot · Independent ML Research · South Kirkby, UK

• HuggingFace: Mati83moni

---

📜 Citation

@article{piesiak2026hybridmoe,
  title   = {HybridMoE Titan v1: A Decoder-Only Language Model Combining 
             Mamba SSM, Multi-Head Attention with RoPE, and Fine-Grained 
             Mixture-of-Experts at 450M Scale},
  author  = {Piesiak, Mateusz},
  year    = {2026},
  note    = {Project Inkblot — Independent ML Research. 
             Available: https://huggingface.co/Mati83moni/HybridMoE-Titan-v1}
}

---

📄 License

CC BY-SA 4.0 — Share and adapt with attribution.

Individual source licenses:

• FineWeb-Edu / FineWeb: ODC-BY (HuggingFace)
• Wikipedia PL: CC BY-SA 3.0
• Wolne Lektury: Public Domain (Polish copyright law)

---

_{Curated with care for independent ML research · 🇵🇱 + 🇬🇧}