---
annotations_creators:
- machine-generated
language_creators:
- expert-generated
language:
- en
license:
- mit
multilinguality:
- monolingual
size_categories:
- n<1K
source_datasets:
- fs90/nano-start-data
task_categories:
- text-generation
pretty_name: Nano-Start Tokenized Dataset
tags:
- educational
- llm-training
- tokenized
- binary
- oxidizr
---

# Nano-Start Tokenized Dataset

Pre-tokenized binary files ready for training with [oxidizr](https://github.com/farhan-syah/oxidizr). This is the tokenized version of [fs90/nano-start-data](https://huggingface.co/datasets/fs90/nano-start-data).

## What is Tokenization?

Language models don't process text directly - they work with numbers called **tokens**. Tokenization converts text into token IDs:

```
"Hello world" → [9906, 1917]
```

This dataset is **pre-tokenized** for simplicity - download and start training immediately. To learn how tokenization works and create your own datasets, see the [splintr](https://github.com/farhan-syah/splintr) project.

## Quick Start

**Option A: Using hf**
```bash
pip install huggingface_hub
hf download fs90/nano-start-data-bin --local-dir data/nano-start/tokenized --repo-type dataset
```

**Option B: Direct download**

Download `combined.bin` from the [Files tab](https://huggingface.co/datasets/fs90/nano-start-data-bin/tree/main) and place it in your project.

**Train with oxidizr:**
```bash
cargo run --release -- \
    --config models/nano-start.yaml \
    --data data/nano-start/tokenized/combined.bin
```

## Files

Download `combined.bin` for training - it contains all data merged together:

| File | Size | Tokens | Description |
|------|------|--------|-------------|
| **`combined.bin`** | 25,516 bytes | 6,379 | **All data merged (recommended)** |

### Individual Files (Optional)

You can also train on individual subsets. Training on different data produces different model behavior:

| File | Size | Tokens | Description |
|------|------|--------|-------------|
| `completions.bin` | 8,788 bytes | 2,197 | Factual statements only |
| `qa.bin` | 11,036 bytes | 2,759 | Q&A pairs only |
| `chat.bin` | 5,692 bytes | 1,423 | Multi-turn conversations only |

Experiment with different files to see how the training data affects model behavior!

## Binary Format

Each `.bin` file contains raw token IDs:

- **Encoding**: u32 (32-bit unsigned integer)
- **Byte order**: Little-endian
- **Headers**: None (raw token stream)
- **Tokenizer**: `cl100k_base` (OpenAI, vocab size: 100,331)

### Reading the Data

```python
import struct

def read_tokens(path):
    with open(path, "rb") as f:
        data = f.read()
    return list(struct.unpack(f"<{len(data)//4}I", data))

tokens = read_tokens("combined.bin")
print(f"Total tokens: {len(tokens)}")
```

## Tokenizer Details

| Property | Value |
|----------|-------|
| Tokenizer | `cl100k_base` (OpenAI GPT-4/GPT-3.5) |
| Vocab size | 100,331 |
| EOS token | `<\|endoftext\|>` (ID: 100257) |

### Special Tokens

| Token | ID | Purpose |
|-------|------|---------|
| `<\|endoftext\|>` | 100257 | Separates examples |
| `<\|system\|>` | 100277 | System instructions |
| `<\|user\|>` | 100278 | User input |
| `<\|assistant\|>` | 100279 | Model response |

## Source Data

To see the human-readable text before tokenization: [fs90/nano-start-data](https://huggingface.co/datasets/fs90/nano-start-data)

## Related Resources

- **Raw data**: [fs90/nano-start-data](https://huggingface.co/datasets/fs90/nano-start-data)
- **Training framework**: [oxidizr](https://github.com/farhan-syah/oxidizr)
- **Tokenization**: [splintr](https://github.com/farhan-syah/splintr) - Learn how to tokenize your own data

## License

MIT License

## Citation

```bibtex
@dataset{nano_start_bin_2024,
  title={Nano-Start Tokenized Dataset},
  author={fs90},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/fs90/nano-start-data-bin}
}
```