Datasets:

cua-lite
/

Multimodal-Mind2Web

license: other
tags:
  - cua-lite
  - gui
  - sft
task_categories:
  - image-text-to-text
configs:
  - config_name: default
    data_files:
      - split: train
        path:
          - '*/*/train*parquet'
          - '*/*/train/*.parquet'
          - '*/*/train/*/*.parquet'
      - split: validation
        path:
          - '*/*/validation*parquet'
          - '*/*/validation/*.parquet'
          - '*/*/validation/*/*.parquet'
  - config_name: web-trajectory
    data_files:
      - split: train
        path:
          - web/trajectory/train*parquet
          - web/trajectory/train/*.parquet
          - web/trajectory/train/*/*.parquet
      - split: validation
        path:
          - web/trajectory/validation*parquet
          - web/trajectory/validation/*.parquet
          - web/trajectory/validation/*/*.parquet

cua-lite/Multimodal-Mind2Web

cua-lite preprocessed version of Multimodal-Mind2Web (osunlp/Multimodal-Mind2Web). Web trajectory data with upstream's canonical four-way split: train + three held-out test sets (test_task, test_website, test_domain) capturing successively harder generalization. The upstream split labels are preserved via metadata.others.split and routed into our validation split; the hash splitter never activates here.

Origin

https://huggingface.co/datasets/osunlp/Multimodal-Mind2Web

Load via `datasets`

from datasets import load_dataset

# entire dataset
ds = load_dataset("cua-lite/Multimodal-Mind2Web")

# just one (platform, task_type) cohort
ds = load_dataset("cua-lite/Multimodal-Mind2Web", "web-trajectory")

You can also filter by metadata.platform / metadata.task_type / metadata.others.* after loading; every row carries a rich metadata struct (see schema below).

Schema

Each row has these columns:

column	type	notes
`image_ids`	list[string]	content-addressed ids (`<sha256>.<ext>`), enables cross-parquet / cross-dataset dedup
`images`	list[Image]	bytes embedded at HF push time; matches `image_ids` index-for-index
`messages`	list[struct]	OpenAI-style turns with `role` + structured `content`
`metadata`	struct	`{platform, task_type, split, others{...}}`

Coordinate values in messages are normalized to [0, 1000] integers.

Layout

<platform>/<task_type>/<split>.parquet                   # single-variant cohort
<platform>/<task_type>/<split>/<variant>.parquet         # multi-variant cohort
<platform>/<task_type>/<split>/shard-NNNNN-of-NNNNN.parquet            # + sharded single-variant
<platform>/<task_type>/<split>/<variant>/shard-NNNNN-of-NNNNN.parquet  # + sharded multi-variant

platform ∈ {desktop, mobile, web}
task_type directory uses a hyphen where the metadata value uses a colon: grounding-action/ → grounding:action
split ∈ {train, validation} — validation is an in-distribution held-out slice (never used in training); test is reserved for out-of-distribution benchmark datasets

Stats

platform	task_type	variant	train	validation
web	trajectory	test_domain	0	478
web	trajectory	test_task	0	99
web	trajectory	test_website	0	83
web	trajectory	train	602	0

Image storage

Images are content-addressed by SHA-256 and deduplicated within this repo. The images column on HuggingFace embeds raw bytes so the Hub viewer renders thumbnails and datasets.load_dataset works out of the box.

For local workflows (SFT export, cross-dataset dedup, split rebalancing), run reverse.py on a cloned repo: it extracts each unique image_id once to a shared image_store/<hash[:2]>/<hash>.<ext> and rewrites the parquets to drop the images column, so rows reference images by hash id only. The shared store is reusable across datasets — the same image in two repos lands in one file.

Total unique images: 7,423
Store size: 7.09 GB

Notes

All three test splits are currently folded into our validation split. A future revision may promote test_website / test_domain to our canonical test split (out-of-distribution benchmark).

License & citation

See original dataset (osunlp/Multimodal-Mind2Web)

See https://huggingface.co/datasets/osunlp/Multimodal-Mind2Web