---
license: mit
language:
- en
pretty_name: a
---
# Auralis-Coder Training Dataset

Packed token dataset for training the **Auralis-Coder** 1.5B parameter language model.

## Dataset Overview

| Property | Value |
|---|---|
| Total tokens | 176,544,387 |
| Tokenizer | Qwen2.5 (Qwen/Qwen2.5-1.5B) |
| Format | uint32 token IDs (packed) |
| Shards | 16 |
| Shard size | ~10M tokens each |
| Documents | 235,219 |
| Duplicates removed | 19 |

## Tokenization & Packing

- Tokenized with Qwen2.5 tokenizer (vocab: 151,936)
- Packed into fixed-size shards (10M tokens each)
- Sequence length: 4,096 tokens during training
- Maximum context: 131,072 tokens

## Usage

```python
import numpy as np

# Load a shard
tokens = np.fromfile("tokens-000000.bin", dtype=np.uint32)

# Reshape into sequences (e.g., 4096 tokens per sequence)
seq_len = 4096
sequences = tokens.reshape(-1, seq_len)
```

## Files

- `tokens-000000.bin` through `tokens-000015.bin` — Packed token shards (uint32)
- `packed_manifest.json` — Dataset statistics and metadata
- `dataset.json` — Additional metadata

## License

Refer to the source dataset licenses (FineWeb, StarCoder, FineWeb-Edu) and the Auralis project distribution.