beomi / megatronlm_dataset_autotokenizer

Megatron-LM/GPT-NeoX compatible Text Encoder with 🤗Transformers AutoTokenizer.

MegatronLM Dataset with 🤗Transformers AutoTokenizer

What's the difference between GPT-NeoX or Megaton-LM?

• GPT-NeoX is a fork of Megatron-LM with some additional features.
• GPT-NeoX supports various tokenizers than Megatron-LM, such as LlamaTokenizer, HFGPT2Tokenizer, or TiktokenTokenizer.

However, both repository does not support Transformers AutoTokenizer out of the box.

That's why I made this repository.

In this repository:

• You can encode your dataset with any AutoTokenizer-compatible tokenizer.

Prepare the dataset

• Checkout examples/sample_dataset.jsonl for the format of the dataset.

// examples/sample_dataset.jsonl
{"text": "This is the first sentence."}
{"text": "This is the second sentence."}
// ...

Prepare the Environment

• Install the dependencies

pip install -r requirements.txt

for more details, please checkout requirements.frozen.txt.

Encode the dataset

• Encode the dataset with the following command
• Check sample shell file:run.sh

# Create output directory
mkdir -p data-sample/
# Encode the dataset
python preprocess_data.py \
--input examples/sample_dataset.jsonl \
--tokenizer-type AutoTokenizer \
--vocab-file beomi/llama-2-ko-7b \
--output-prefix data-sample/out \
--dataset-impl mmap \
--workers 10

Notes

1. Input file name must ends with .jsonl

2. Your JSONL file must NOT contain any empty lines.

Example(This makes error):

{"text": "This is the first sentence."}
{"text": "This is the second sentence."}
// YOU SHOULD NOT HAVE EMPTY LINE HERE
{"text": "This is the third sentence."}
// YOU SHOULD NOT HAVE EMPTY LINE AT THE END OF THE FILE TOO

JSON above will cause this error:

jsonlines.jsonlines.InvalidLineError: line contains invalid json: Expected object or value (line 3)

This is OK:

{"text": "This is the first sentence."}
{"text": "This is the second sentence."}

Acknowledgement

This repo codes are partially copied/edited from EleutherAI/gpt-neox @ 10bf78.