README.md

---
dataset_info:
  features:
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 4483331
    num_examples: 342
  - name: validation
    num_bytes: 622617
    num_examples: 39
  download_size: 2534957
  dataset_size: 5105948
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
license: apache-2.0
---
# Q Code Pretraining Corpus

This dataset provides a corpus of Q programming language code and documentation, curated for pretraining large language models and code models.

## 📊 Dataset Overview

- **Total Data**: Over 1.6 million Q tokens, 5+ million characters
- **Documents**: 342 training chunks, 39 validation chunks
- **Source Types**:  
  - Open-source Q repositories (MIT/Apache 2.0 licenses)  
  - Official KDB+/Q documentation and tutorials  
  - Hand-curated code snippets and scripts
- **Format**: Cleaned, deduplicated, chunked for efficient pretraining

## 🎯 Key Features

- **Q-Only**: All data is pure Q language (no mixed Python or non-code noise)
- **Permissive Licensing**: All source code is MIT or Apache 2.0, suitable for both research and commercial use
- **Coverage**: Includes code from analytics, time-series, database queries, and utilities
- **Filtered & Scored**: LLM-assisted quality scoring plus manual review for top-tier data fidelity
- **Chunked & Ready**: Delivered as 4k-token chunks for immediate use with Hugging Face, TRL, or custom pipelines

## 🏗️ Dataset Structure

Each record is a text chunk, containing code or documentation in Q.

Splits:
- `train`: Main corpus for pretraining (342 chunks)
- `validation`: Holdout set for evaluation (39 chunks)

Sample record:
```python
{
    "text": str   # Raw Q code or documentation chunk
}
```

## 🧑‍💻 Usage

### Loading the Dataset

```python
from datasets import load_dataset

# Load the full Q pretraining dataset
dataset = load_dataset("morganstanley/q_pretrained_dataset")

# Access splits
train_data = dataset["train"]
val_data = dataset["validation"]
```

### Example: Previewing Data

```python
sample = dataset["train"][0]
print(sample["text"])
```

### Training Usage

This dataset is designed for language model pretraining using next-token prediction or masked language modeling objectives.  
Supports efficient training with Hugging Face Transformers, TRL, or custom frameworks.

## 🔤 About Q Programming Language

Q is a vector and array programming language developed by Kx Systems for high-performance analytics, finance, and time-series applications.

It features:
- Concise, functional, array-oriented syntax
- Powerful built-in operators for large-scale data manipulation
- Industry adoption in trading, banking, and real-time analytics

## 📁 Source Repositories

Major open-source Q repos included:
- DataIntellectTech/TorQ
- psaris/qtips
- psaris/funq
- KxSystems/ml
- finos/kdb
- LeslieGoldsmith/qprof
- jonathonmcmurray/reQ
- ...and more

All with permissive licenses (MIT or Apache 2.0).

## 📈 Data Preparation & Filtering

- **Automated Scoring**: Qwen-2.5-32B was used to score each file (0–10) for quality and relevance; only files scoring ≥4 were included.
- **Manual Review**: Additional cleaning to remove non-Q files or low-value content.
- **Deduplication**: Duplicate and boilerplate code removed.

## 📝 Citation

If you use this dataset in your research, please cite:

```bibtex
@dataset{q_pretraining_corpus_2024,
    title={Q Code Pretraining Corpus},
    author={Brendan Rappazzo Hogan},
    year={2024},
    url={https://huggingface.co/datasets/bhogan/q-pretraining-corpus},
    note={Dataset for domain-adaptive pretraining of language models on the Q programming language}
}
```

**Associated Paper:** [Link to paper will be added here]