Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Curriculum Learning Dataset - 300M Tokens

This dataset is prepared for curriculum learning with separate source files for dynamic mixing during training.

Dataset Information

Total Tokens: 358,000,524
Total Documents: 1,594,150
Tokenizer: tiktoken cl100k_base (GPT-4)
Vocabulary Size: 100,277

Sources

Wikitext

Tokens: 102,000,024 (28.5%)
Documents: 700,699
File: wikitext_train.bin

Ultrachat

Tokens: 208,000,134 (58.1%)
Documents: 857,922
File: ultrachat_train.bin

Stack

Tokens: 48,000,366 (13.4%)
Documents: 35,529
File: stack_train.bin

Files

wikitext_train.bin - WikiText-103 training data (facts and encyclopedic knowledge)
stack_train.bin - The Stack training data (code and logical reasoning)
ultrachat_train.bin - UltraChat training data (conversational behavior)
val.bin - Validation set (mixed, fixed distribution)
dataset_stats.json - Dataset statistics and metadata

Curriculum Learning Strategy

This dataset enables curriculum learning with progressive distribution changes across epochs:

Phase	Epoch	WikiText (Facts)	Stack (Logic)	UltraChat (Behavior)	Objective
Current	1	28%	13%	59%	Pattern Discovery
Foundation	2-3	70%	20%	10%	Knowledge/Logic Hardening
Bridge A	4	50%	25%	25%	Balanced Contextualization
Bridge B	5	35%	25%	40%	Priority Shift
Bridge C	6	20%	20%	60%	Instruction Emergence
Refinement	7-8	10%	10%	80%	Behavior & Formatting

Usage

Automatic Download (Recommended)

The dataset will be automatically downloaded when you start training:

python train.py --config config/gpu_training_117m_balanced.yaml

Configuration

In your training config:

data:
  dataset_type: "binary"
  train_file: "data/balanced_300m_curriculum/train.bin"
  val_file: "data/balanced_300m_curriculum/val.bin"
  
  huggingface_dataset:
    repo_id: "0x-genesys/mix_wiki_code_chat_data_300M_tokens_curriculum"
    dataset_name: "balanced_300m_curriculum"
    auto_download: true

training:
  curriculum_learning:
    enabled: true
    sources:
      - "wikitext"
      - "stack"
      - "ultrachat"
    epoch_distributions:
      1:  [28, 13, 59]
      2:  [70, 20, 10]
      3:  [70, 20, 10]
      4:  [50, 25, 25]
      5:  [35, 25, 40]
      6:  [20, 20, 60]
      7:  [10, 10, 80]
      8:  [10, 10, 80]

Manual Download

from huggingface_hub import hf_hub_download

# Download all source files
for source in ['wikitext', 'stack', 'ultrachat']:
    hf_hub_download(
        repo_id="0x-genesys/mix_wiki_code_chat_data_300M_tokens_curriculum",
        filename=f"{source}_train.bin",
        repo_type="dataset",
        local_dir="data/balanced_300m_curriculum"
    )

# Download validation file
hf_hub_download(
    repo_id="0x-genesys/mix_wiki_code_chat_data_300M_tokens_curriculum",
    filename="val.bin",
    repo_type="dataset",
    local_dir="data/balanced_300m_curriculum"
)

Citation

If you use this dataset, please cite:

@dataset{{curriculum_300m_2024,
  title={{Curriculum Learning Dataset - 300M Tokens}},
  author={{0x-genesys}},
  year={{2024}},
  publisher={{Hugging Face}},
  howpublished={{\url{{https://huggingface.co/datasets/0x-genesys/mix_wiki_code_chat_data_300M_tokens_curriculum}}}}
}}

License

This dataset combines data from multiple sources. Please refer to the original licenses:

WikiText-103: Creative Commons Attribution-ShareAlike License
The Stack: Multiple licenses (see The Stack dataset)
UltraChat: MIT License

Model Repository: 0x-genesys/neo_weights_checkpoints
Combined Dataset: 0x-genesys/mix_wiki_code_chat_data_300M_tokens
Training Code: GitHub Repository

Downloads last month: 28

Total file size:

1.43 GB

Datasets:

0x-genesys
/

mix_wiki_code_chat_data_300M_tokens_curriculum