README.md

---
license: apache-2.0
language:
- en
- ru
- code
task_categories:
- text-generation
- text2text-generation
- question-answering
size_categories:
- 100B<n<1T
tags:
- multimodal
- code
- math
- pre-training
- instruction-tuning
---

# Dataset Card for LLM_multimodal

**LLM_multimodal** is the official training corpus designed for the **LLM D6** model series. It contains a massive, high-quality collection of **300 Billion tokens**, carefully curated to balance linguistic diversity, mathematical reasoning, and programming capabilities. 

This repository hosts both the raw/processed pre-training data and the instruction-following datasets used for supervised fine-tuning (SFT).

# tokenizer is located at LLM_D6 in my models!
---


## 📊 Dataset Summary

* **Total Token Count:** ~300 Billion tokens
* **Languages:** English (`en`), Russian (`ru`), and various Programming Languages.
* **Domains:** General Web Text, Academic/Mathematical Papers, Code Repositories, and Conversational data.
* **Intended Model:** LLaMA-3 architectures (specifically the D6 0.8B parameter custom build).
* **Tokenizer Compatibility:** This dataset was used to train a custom tokenizer with a **52,000 vocabulary size**, optimized for English, Cyrillic (Russian) scripts, and programming syntax.

---

## 📂 Dataset Structure

The dataset is divided into two primary subsets to support the full lifecycle of Large Language Model training:

### 1. `pre-train` (Unsupervised Corpus)
This subset contains the bulk of the 300B tokens used for the foundational training phase. It is structured as large text documents (typically stored in `.parquet` or `.jsonl` formats) with a focus on high-quality, deduplicated data.

* **English & Russian Text:** Filtered web crawls, encyclopedias, and literature to ensure strong bilingual fluency.
* **Code:** Source code from permissive repositories covering Python, C++, JavaScript, and other major languages.
* **Math:** LaTeX-formatted equations, mathematical proofs, and STEM-focused academic texts to enhance reasoning.

### 2. `fine-tune` (Instruction/SFT Corpus)
A curated subset of high-quality prompt-completion pairs formatted for dialogue and instruction following. 
* **ChatML Format Ready:** The data structure natively aligns with standard ChatML prompt templates (`<|bos|>`, `<|end_of_text|>`, role definitions).
* **Tasks:** Includes physical commonsense reasoning (PIQA), general reasoning (HellaSwag), multi-task accuracy (MMLU), and Russian language understanding (XWinograd RU).

---

## 🚀 Usage

You can easily load this dataset using the `datasets` library from Hugging Face:

```python
from datasets import load_dataset

# Load the fine-tuning dataset
sft_dataset = load_dataset("firdavsus/LLM_multimodal", split="fine_tune")

# Load the pre-training dataset (Requires streaming due to size)
pretrain_dataset = load_dataset("firdavsus/LLM_multimodal", split="pre_train", streaming=True)

# Example output from the SFT dataset
print(sft_dataset[0])