akshat-OwO/nev-bench

Nev browser-step benchmark

Test data for browser decision models such as Nev-0.8B. Each record is one browser step, written as a System One request with a label on every question. A model reads the page as numbered elements and answers four typed questions: is the step done, which operation comes next, which element it acts on, and which input value goes in a field.

The scripts that built this data, trained Nev, and score models are in scripts/. Full results, including other models, are in docs/research/nev.md of the Contingency repository.

Contents

Path	Records	Questions	What it is
heldout/test.jsonl	550	1,488	Steps on sites never used in training: 250 from the Contingency crawl, 300 from WebChain
heldout/calibration.jsonl	332	951	Steps on other held-out sites, for fitting a model's temperature
mind2web-ids/*.txt	400 each	–	Mind2Web test steps that were scored, as action_uid lists

Mind2Web asks users not to publish its test set in plain text, to prevent leakage into training data. This dataset therefore holds only the IDs of the scored Mind2Web steps. To rebuild those records from the official Multimodal-Mind2Web files, run scripts/m2w_build.py. The rebuilt records match the scored ones exactly.

Record format

{
  "state": {
    "elements": [{"index": "1", "role": "combobox", "name": "City"}, "..."],
    "page": {"content": ["heading: Delivery location", "..."], "title": "Shop", "url": "https://shop.example/"},
    "recent_actions": ["FILL [6] textbox \"Search\" ← \"Sector 14\" → effect: dom"]
  },
  "questions": {
    "done":      {"type": "noul",   "instructions": {"condition": "...", "question": "..."}, "criteria": {...}, "label": false},
    "operation": {"type": "choice", "instructions": {"step": "...", "question": "..."}, "criteria": {...}, "label": "CLICK"},
    "click_target": {"type": "choice", "instructions": {...}, "criteria": {"2": "[2] button \"Save location\""}, "label": "2"}
  },
  "_meta": {"source": "webchain", "website": "example.com"}
}

A record asks only the questions that have a defined answer. For a choice question the label is the option name. For a noul question it is true or false.

Scoring

"Steps fully right" counts a step only when its operation and its *_target answers are both correct. Report it alongside accuracy per question kind (done, operation, target, value).

Model	Steps fully right on heldout/test
Nev-0.8B	76.9%
Laya (multilingual, 8,192 tokens)	19.7%

Scripts

Script	Purpose
score_endpoint.py	Score any System One endpoint (Nev or Kev through kev.serve, Jev, Clef) on these files
m2w_build.py, m2w_convert.py	Rebuild the Mind2Web test files from the official source
gen.py	Turn crawled pages into labelled requests, split by site
webchain_convert.py, data_modal.py	Convert WebChain and Mind2Web training data on Modal CPUs
kev_modal.py, launch_train.py, launch_score.py	Train and score on Modal, detached (adapted from Kev, Apache-2.0)
laya_modal.py	Score Laya on the same files
mlx_train.py, score.py, regress.py	Score and fit a temperature locally on Apple Silicon
sites.txt	The 173 start URLs of the crawl

The crawler and the live Flow Skill benchmark are TypeScript and live in the Contingency repository, in packages/cli/tests/benchmarks.

Score Nev on this benchmark

These steps need uv and a clone of Kev with uv sync --extra serve.

1. Start Nev in the Kev checkout:

   uv run --extra serve python -m kev.serve --run akshat-OwO/nev-0.8b --port 8009
   

2. Download this dataset and score the held-out set:

   hf download akshat-OwO/nev-bench --type dataset --local-dir nev-bench && cd nev-bench
   python3 scripts/score_endpoint.py heldout/test.jsonl --sample 300 --model nev
   

3. To score Mind2Web, rebuild a test split, then score it by ID:

   uv run --with lxml --with cssselect --with pyarrow --with "huggingface_hub[hf_xet]" \
     python scripts/m2w_build.py --split test_website --out data/m2w
   python3 scripts/score_endpoint.py data/m2w/test_website.explicit.jsonl \
     --ids mind2web-ids/m2w_test_website.explicit.txt --sample 300 --model nev
   

To score another endpoint, set JEV_URL, and JEV_API_KEY if it needs one, and pass its model name with --model.

Sources and license

• Contingency crawl: browser snapshots of public websites captured with Contingency, with instructions and labels generated from the recorded page and action.
• WebChain (CC BY 4.0): human trajectories from webagentlab/webchain, converted to this format. The 26 sites here were held out from Nev's training.

This dataset is released under CC BY 4.0. Credit WebChain and this dataset when you use it.