| --- |
| 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. |
