dataset_info:
features:
- name: text
dtype: string
splits:
- name: train
num_bytes: 0
num_examples: 1223305
download_size: 3035467525
dataset_size: 0
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
license: mit
language:
- en
- ta
size_categories:
- 1M<n<10M
🚀 ZFusionAI Pretraining Dataset
Welcome to the ZFusionAI repository! This dataset is a carefully curated, high-quality collection of text data designed specifically for training small-to-medium-sized language models (in the 500M to 1B parameter range).
By combining general web knowledge, specialized mathematical reasoning, coding proficiency, and Indic language capabilities, this dataset provides a balanced foundation for versatile model performance.
📊 Dataset Composition
We have synthesized four distinct data sources to ensure a balanced distribution of knowledge. The data is processed via weighted interleaving to maintain high structural diversity.
| Data Source | Domain | Weight | Purpose |
|---|---|---|---|
| FineWeb-Edu | General Web / Edu | 30% | General world knowledge & language fluency |
| UltraData-Math | Mathematics | 20% | Logical reasoning & symbolic problem solving |
| Python-Github | Programming | 20% | Syntax, logic, and functional code generation |
| IndicCorpV2 | Tamil (ta) | 30% | Multilingual support & linguistic diversity |
🛠️ Data Processing Pipeline
To ensure the model learns from high-quality tokens, we implemented a rigorous filtering process:
- Streaming Load: Efficient memory management by processing data in streams. 🌊
- Normalization: Harmonized schema (all columns mapped to
text). 🔄 - Length Filtering: Dropped samples with fewer than 30 words to avoid noise. ✂️
- Quality Heuristics: Applied an alphabet/number ratio check (min 30% density) to filter out "junk" characters, HTML boilerplate, and non-informative strings. 🧹
⚙️ How to use
You can load this dataset easily using the Hugging Face datasets library:
from datasets import load_dataset
# Load the ZFusionAI dataset
dataset = load_dataset("GTKING/ZFusionAI_Pretraining_Data", streaming=True)
# Peek at the first example
print(next(iter(dataset)))
💡 Best Practices for 500M-1B Models
Given the parameter count, this dataset is optimized to encourage data efficiency:
- Tokenizer: Ensure you are using a tokenizer compatible with the mixed content (e.g., a byte-level BPE tokenizer) to handle both code and Tamil characters effectively. 🧩
- Epochs: With a dataset of this scale, consider 1–3 epochs depending on your compute budget. ⏳
- Instruction Tuning: While this is a pre-training corpus, it is highly recommended to follow up with an instruction-tuning phase using specialized datasets for optimal chat performance. 🤖
⚖️ License & Credits
- FineWeb-Edu: HuggingFaceFW
- UltraData-Math: OpenBMB
- IndicCorpV2: AI4Bharat
- Python-Github: Sourced via Hugging Face
This dataset is provided for research and development purposes. 🎓