pszemraj/text2image-multi-prompt

text2image multi-prompt(s): a dataset collection

• collection of several text2image prompt datasets
• data was cleaned/normalized with the goal of removing "model specific APIs" like the "--ar" for Midjourney and so on
• data de-duplicated on a basic level: exactly duplicate prompts were dropped (after cleaning and normalization)

updates

• Oct 2023: the default config has been updated with better deduplication. It was deduplicated with minhash (params: n-gram size set to 3, deduplication threshold at 0.6, hash function chosen as xxh3 with 32-bit hash bits, and 128 permutations with a batch size of 10,000.) which drops 2+ million rows.
  • original version is still available under config_name="original"

contents

default:

DatasetDict({
    train: Dataset({
        features: ['text', 'src_dataset'],
        num_rows: 1677221
    })
    test: Dataset({
        features: ['text', 'src_dataset'],
        num_rows: 292876
    })
})

For original config:

DatasetDict({
    train: Dataset({
        features: ['text', 'src_dataset'],
        num_rows: 3551734
    })
    test: Dataset({
        features: ['text', 'src_dataset'],
        num_rows: 399393
    })
})

NOTE: as the other two datasets did not have a validation split, the validation split of succinctly/midjourney-prompts was merged into train.