metadata
configs:
- config_name: default
data_files:
- split: train
path: train-*.parquet
- split: val
path: val-*.parquet
dataset_info:
features:
- name: image
dtype: image
- name: caption
dtype: string
- name: source
dtype: string
- name: split
dtype: string
- name: image_id
dtype: string
- name: source_index
dtype: int64
- name: source_dataset
dtype: string
- name: source_split
dtype: string
- name: source_url
dtype: string
- name: filename
dtype: string
- name: width
dtype: int32
- name: height
dtype: int32
- name: image_sha256
dtype: string
- name: caption_index
dtype: int32
WebMix Parquet
Image-text parquet dataset generated from: cc12m, datacomp12m, yfcc15m, mscoco.
Images are stored in the Hugging Face Image column format as encoded bytes.
Training rows come from --sources; validation/support rows come from
--eval-sources.
Generation arguments:
{
"cache_dir": null,
"chunk_size": 10000,
"compression": "zstd",
"download_max_retries": 5,
"download_num_proc": 8,
"eval_sources": [
"mscoco"
],
"hf_num_proc": 1,
"image_format": "jpeg",
"jpeg_quality": 90,
"limit_per_source": null,
"log_every": 10000,
"min_image_size": 16,
"multiprocessing_context": "fork",
"num_proc": 16,
"out_dir": "data/webmix_parquet",
"overwrite": false,
"resume": true,
"rows_per_shard": 100000,
"sources": [
"cc12m",
"datacomp12m",
"yfcc15m"
],
"url_concurrency": 16,
"url_retries": 4,
"url_timeout": 10,
"worker_maxtasksperchild": 100
}