README.md

metadata

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

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.