webbrain-vl-2-450M-dataset

50,000 browser viewport screenshots, each paired with the six-section observation that WebBrain's vision subsystem produces for its planning agent: what the page is, the exact visible text, the inputs, the state signals, the blockers and the unknowns. Labels were generated by a teacher VLM (Qwen3.6-35B-A3B) with DOM/accessibility evidence, then filtered for structure and grounded exact text.

It is a screenshot-to-description dataset. It contains no next actions and no click coordinates.

This dataset trains webbrain-one/webbrain-vl-2-450M. The browser-ready export is webbrain-one/webbrain-vl-2-450M-onnx.

At a glance

Rows	46,879 train + 3,121 validation (50,000 unique screenshots)
Modality	One screenshot in, one six-section text observation out
Format	Parquet with embedded images (data/*.parquet, 3.74 GB)
Data frozen	2026-08-22 (data commit fb191a0b73e0906e6c3b9f35f6ec0dba1337b72c); no rows added since
Teacher	qwen/qwen3.6-35b-a3b (49,634 labels) and Qwen/Qwen3.6-35B-A3B-Q4_K_M (366)
Built on	The 16,646-row WebBrain VL-1 set plus 33,354 targeted synthetic browser screens
Reused by	WebBrain's VL-3 and VL-4 training mixtures, verbatim (see Lineage)
License	Per-row upstream licenses: MIT or CC-BY-4.0 (see Sources and licenses)

Data cutoff

All 50,000 rows were frozen before the 2026-08-22 upload, and later commits changed only this card and its sample image. Each row reflects its upstream snapshot rather than the live web:

Source	Snapshot	Snapshot date
HuggingFaceM4/WebSight (config v0.1)	revision b11f8172f89c992b56ac702319e02c428cca4a4e	2024-03-26
docling-project/screenparse	revision bc0113f94e262ac39bf881c60547a09b5b4e4462	2026-05-22
Farama-Foundation/miniwob-plusplus	tasks rendered locally through ServiceNow/BrowserGym, seeded	rendered for this release
webbrain-one/targeted-browser-gui-synthetic	template-generated pages for this release	generated for this release

Nothing in the dataset is newer than 2026-08-22.

What the target looks like

Every row uses the same production system prompt and user request. The response is terse and always has these six numbered sections:

1. Page purpose: one line, for example "Gmail compose" or "Stripe checkout form".
2. Visible text: the exact strings on buttons, links, headings, tabs and menu items, quoted verbatim.
3. Inputs: each visible form field with its label, placeholder, current value and focus/disabled state.
4. State signals: spinners, toasts, modals, error banners, success messages, CAPTCHAs, consent banners, overlays.
5. Blockers: anything that would stop the next likely action (overlay, disabled submit, missing data, auth prompt).
6. Unknowns: anything not clearly readable. The model must not guess numbers, names or identifiers.

Responses have a median of 677 characters (95th percentile 965, maximum 2,702).

Samples

A validation row from the targeted synthetic set (category modal_consent, locale tr), shown in full except the shared prompts:

Field	Value
id	targeted-v2-modal_consent-023500-5d2b4f27
group_id	targeted-v2-modal_consent-00470
source_dataset / source_license	webbrain-one/targeted-browser-gui-synthetic / CC-BY-4.0
source_metadata_json	{"category": "modal_consent", "locale": "tr", "template": "modal_case", "synthetic": true, ...}
teacher_model	qwen/qwen3.6-35b-a3b
image	1366x768 PNG (above)

response:

1) Page purpose: Northstar project settings page blocked by a cookie consent dialog.
2) Visible text:
- "Northstar"
- "Workspace / Project settings"
- "TR"
- "View WB-023500"
- "Overview"
- "Current task"
- "Activity"
- "Settings"
- "Project settings"
- "Profile"
- "Members"
- "Billing"
- "Çerez tercihleri"
- "Analytics is optional for workspace WB-04500. Essential storage is always active."
- "İsteğe bağlıları reddet"
- "Seçilenleri kabul et"
3) Inputs: None visible.
4) State signals: A modal dialog titled "Çerez tercihleri" is present, blocking interaction with the main page content.
5) Blockers: The cookie consent modal is blocking the main viewport.
6) Unknowns: None.

messages_json wraps the same row as a three-turn conversation:

[
  {"role": "system", "content": [{"type": "text", "text": "You are the vision subsystem of a web-automation agent. A screenshot of the current browser viewport is attached. Describe what is on screen so the planning agent can decide its next action.\n\nFormat — keep it terse, structured, no flowery prose: ..."}]},
  {"role": "user", "content": [{"type": "image", "image": "images/<image_sha256>.png"},
                               {"type": "text", "text": "Describe this screenshot of the current browser viewport for a web-automation agent. Follow the format in the system prompt."}]},
  {"role": "assistant", "content": [{"type": "text", "text": "1) Page purpose: Northstar project settings page blocked by a cookie consent dialog.\n2) Visible text: ..."}]}
]

The other sources look the same from the outside. A WebSight row, for comparison, opens with "1) Page purpose: Travel agency homepage with destination search and category navigation." and lists the search box under Inputs; a MiniWoB++ row opens with "1) Page purpose: MiniWoB browser task: phone-book ...". Browse more rows in the dataset viewer above.

Schema

• id: row identifier
• image: browser viewport screenshot
• system_prompt: WebBrain production vision system prompt (identical in every row)
• user_prompt: WebBrain production screenshot request (identical in every row)
• response: accepted six-section teacher label
• messages_json: leap-finetune-compatible multimodal conversation (system, user with image, assistant)
• group_id: split/leakage group
• image_sha256: image identity
• source_*: source, license, URL, and provenance (source_metadata_json holds source-specific fields such as WebSight/ScreenParse revision and sample id, the MiniWoB task and seed, or the synthetic category and locale)
• teacher_model: teacher identifier; private endpoint details are not included

Splits

• train: 46,879
• validation: 3,121

Splits are grouped by task/site (12,983 group_id values) and deduplicated by image SHA-256. WebBrain's 100-case test/vision suite is excluded and remains held out: the release manifest records zero exact and zero perceptual overlap with it.

Source	Train	Validation
webbrain-one/targeted-browser-gui-synthetic	31,314	2,040
HuggingFaceM4/WebSight	9,069	492
Farama-Foundation/miniwob-plusplus	4,127	465
docling-project/screenparse	2,369	124

Sources and licenses

Source	License	Rows
Farama-Foundation/miniwob-plusplus	MIT	4,592
HuggingFaceM4/WebSight	CC-BY-4.0	9,561
docling-project/screenparse	CC-BY-4.0	2,493
webbrain-one/targeted-browser-gui-synthetic	CC-BY-4.0	33,354

This combined dataset uses license: other because source licenses differ. Users must comply with every row's source_license and the linked upstream terms. WebSight use must be disclosed and attributed under CC-BY-4.0. MiniWoB++ is MIT. ScreenParse declares CC-BY-4.0; preserve its attribution and review sampled third-party page content. Project-generated synthetic GUI rows are CC-BY-4.0. Do not assume project-owned or authorized captures grant rights beyond their recorded license.

What the synthetic rows cover

The 33,354 synthetic screens target capabilities the legacy sources lacked:

Category	Rows	Category	Rows
multilingual OCR	8,000	auth blockers	2,500
form validation	4,500	calendar / kanban / map	2,500
data tables	3,500	dashboards / charts	2,500
modals and consent	3,500	loading states and toasts	2,500
general browser pages	2,354	occlusion and uncertainty	1,500

Synthetic page locales: English 8,535, Turkish 5,168, German 3,848, French 3,657, Spanish 3,643, Japanese 2,956, Arabic 2,901, Korean 2,646. Labels are written in English and quote on-screen text verbatim in its original language. The MiniWoB++ rows cover 29 tasks (click, tabs, tables, forms, navigation trees and similar).

Teacher

• Qwen/Qwen3.6-35B-A3B-Q4_K_M: 366 accepted labels
• qwen/qwen3.6-35b-a3b: 49,634 accepted labels

Teacher outputs can contain OCR errors or hallucinations. The automated quality gate does not replace manual review. Do not use this dataset for security, payment, identity, or accessibility decisions without independent validation.

Lineage

webbrain-vl-1-450M-dataset   (private, 2026-08-21)  16,646 rows: MiniWoB++ 4,592, WebSight 9,561, ScreenParse 2,493
  └─ webbrain-vl-2-450M-dataset   (this repo, 2026-08-22)  those 16,646 legacy rows + 33,354 targeted synthetic = 50,000
       └─ WebBrain VL-3 dataset   (private, 2026-09-07)  this repo copied verbatim (pinned at 5403825) as its `vl2` source,
            │                                            plus curated real-site screenshots, MolmoPoint-GUISyn grounding and WebChain
            └─ WebBrain VL-4 dataset   (private draft, 2026-09-27)  an unchanged copy of VL-3

Models trained on it:

Model	How this dataset was used
webbrain-one/webbrain-vl-2-450M (+ -onnx)	All 46,879 train rows, one epoch (LFM2.5-VL-450M LoRA)
WebBrain VL-3 450M (private)	All rows, up to two passes, as 61,234 of 335,572 VL-3 training draws
webbrain-one/minicpmv47-webbrain-v1	Same 61,234 draws, as part of VL-4
webbrain-one/webbrain-compass-vld-tiny-v5	12,000-draw replay of these rows in round 2

Usage

The splits load directly with 🤗 Datasets, and the image column decodes to a PIL image:

import json
from datasets import load_dataset

ds = load_dataset("webbrain-one/webbrain-vl-2-450M-dataset", split="train")
row = ds[0]
row["image"]               # PIL.Image
print(row["response"])     # six-section observation

# messages_json references "images/<sha>.png"; swap in the decoded screenshot to get a ready chat sample
def to_conversation(row):
    messages = json.loads(row["messages_json"])
    for message in messages:
        for part in message["content"]:
            if part.get("type") == "image":
                part["image"] = row["image"]
    return messages

Use streaming=True to inspect rows without downloading all 3.7 GB. Pin revision="fb191a0b73e0906e6c3b9f35f6ec0dba1337b72c" (or any later commit; the data files are unchanged) for reproducibility.

To recreate local image files plus a typed messages JSONL for Liquid AI's leap-finetune (what this project's scripts/materialize_hf_dataset.py utility does), write each image to the path its messages_json already references:

import json
from pathlib import Path
from datasets import load_dataset

out = Path("vl2-materialized")
(out / "images").mkdir(parents=True, exist_ok=True)
for split in ("train", "validation"):
    ds = load_dataset("webbrain-one/webbrain-vl-2-450M-dataset", split=split)
    with open(out / f"{split}.jsonl", "w", encoding="utf-8") as fh:
        for row in ds:
            messages = json.loads(row["messages_json"])
            image_path = next(p["image"] for m in messages for p in m["content"] if p.get("type") == "image")
            row["image"].save(out / image_path)
            fh.write(json.dumps({"messages": messages}, ensure_ascii=False) + "\n")

Fine-tuning

The released model's recipe (from the webbrain-vl-2-450M card):

Item	Value
Base model	LiquidAI/LFM2.5-VL-450M
Method	Supervised LoRA fine-tuning, then merge
LoRA	rank 16, alpha 32, dropout 0.05
Schedule	one epoch, 2,929 steps, effective batch 16, learning rate 1e-4, cosine, 3% warmup
Images	max_image_tokens 256 (the LFM2.5-VL processor default)
Precision / hardware	BF16 with gradient checkpointing, one RTX 4090 24 GB
Selection	validation loss on the 3,121-row validation split; the 100-case suite was never used for selection

Train on the assistant turn only. With the Transformers processor this means masking every token before the response:

from transformers import AutoProcessor

processor = AutoProcessor.from_pretrained("LiquidAI/LFM2.5-VL-450M", revision="fc6221ca597f3315e4f82fc2df606783267b34ba")

def encode(row):
    conversation = to_conversation(row)
    prompt = processor.apply_chat_template(conversation[:-1], add_generation_prompt=True, tokenize=False)
    full = processor.apply_chat_template(conversation, tokenize=False)
    batch = processor(text=[full], images=[[row["image"]]], return_tensors="pt")
    target = processor.tokenizer(full[len(prompt):], add_special_tokens=False)["input_ids"]
    labels = batch["input_ids"].clone()
    labels[:, : labels.shape[1] - len(target)] = -100   # loss only on the six-section response
    batch["labels"] = labels
    return batch

The messages_json column is already in the leap-finetune conversation format, so the materialized JSONL above can be passed to leap-finetune's VLM SFT configs without conversion. Evaluate with WebBrain's public 100-case test/vision suite, which this dataset excludes.

Intended use

Supervised fine-tuning and evaluation of compact VLMs that turn browser screenshots into structured observations for a downstream automation planner. It is not a next-action or click-coordinate dataset.

Integrity

release-manifest.json lists every shard with its row count, byte size and SHA-256, the locked processed snapshot ID (fb55b3afe2026634d5fadbe35dc4bae5b00602b44634459f65ca93e27f227218) and the held-out overlap check. SHA256SUMS lists the SHA-256 of every data shard and the release manifest; its README.md entry does not track later card edits.

Changelog

• 2026-08-22: data uploaded (fb191a0), card links refreshed (5403825).
• 2026-10-30: card expanded with data cutoff, samples, lineage, usage and fine-tuning sections; the sample screenshot (a copy of one validation image) added under assets/. Data files unchanged.