Dataset versions
LexBench-Web-Fetch uses vMAJOR.MINOR.PATCH tags on Hugging Face. Each tag names
one immutable dataset commit. Keep the resolved full commit, task and reference
hashes, Browser Eval/scorer commit, and execution settings with every experiment.
Dataset tags are independent of Browser Eval's software version.
Release populations
| Dataset edition | Tasks | Scope | Provenance |
|---|---|---|---|
| v1.0.0 | 430 | Earlier population: 100 each for general web, dynamic sites, popular sites and access restrictions, plus 10 embedded documents, 17 media-information tasks and 3 direct-media tasks | The four original suites' task, catalog and source-input hashes match bench-v1.0.0; seven viewer groups select the 430-task population |
| v2.1.0 | 550 | Expanded catalog: the 430-task population, 100 article tasks and 20 additional dynamic/challenge examples | Uses the existing 2.1.0 catalog declaration and reviewed reference/scoring provenance, following the earlier bench-v2.0.1 snapshot |
The 550-task catalog follows the earlier 430-task population. Adding tasks
starts a new major series. A later refresh of the 430-task viewer does not make
that population the newer benchmark. Retained historical definitions under
definitions/ do not expand the selected population. The viewer configurations and their
hash-bound JSONL files identify the active task entries. A task is identified by
its source suite and task ID; counts are task entries, not unique domains or URLs.
The former repository used bench-v1.0.0 for a historical 430-task report and
bench-v2.0.1 for the expanded 550-task catalog. Those GitHub tags identify their
original code and catalogs; HF tags select the active dataset population.
The later 550-task snapshot already declares catalog version 2.1.0. It is not
relabeled as 2.0.1. Original tags and commits remain available as provenance.
Matching labels or counts do not certify an old report's scorer or measurements.
Historical reports keep their original task, reference and scoring hashes.
Version increments and comparisons
| Change | Increment | Comparing results |
|---|---|---|
| Add/remove tasks, or change their URLs, requests, required output scope or permitted retrieval routes | Major | Report full-population results separately |
| Correct references or acceptance rules while preserving the original task inputs and intended measurement | Minor | Rescore valid saved outputs from both cohorts under the same corrected rules |
| Correct documentation, metadata or presentation without changing tasks, references or scoring | Patch | Direct comparison when execution settings also match |
Routine commits do not require a new release. Prepare changes through a Hugging Face pull request, validate the intended release snapshot, then publish a tag after explicit release authorization. Never move or overwrite a published tag. Keep historical snapshots available. A dataset release does not certify a new scorer or change an existing report's recorded edition.
Pin an experiment
Resolve a readable tag to a full commit before loading data:
from datasets import load_dataset
from huggingface_hub import HfApi
repository = "Lexmount/LexBench-Web-Fetch"
revision = HfApi().dataset_info(repository, revision="v1.0.0").sha
tasks = load_dataset(repository, "general-web", split="test", revision=revision)
Save revision in the experiment manifest. Select every configuration needed by
the declared population; one configuration is not the complete benchmark.
manifest.json records file hashes and the edition's active population. The
original catalog under definitions/benchmark-release.json retains its own
historical scorer identity and must not be relabeled as the new dataset release.
The edition's population_sha256 hashes the sorted list of suite, id and
task_sha256 records, ordered by suite and ID. Each task and the list use UTF-8
JSON with ensure_ascii=False, sorted keys, two-space indentation and a final
newline before SHA-256 hashing. Viewer JSONL file hashes remain in files.
Browser Eval selects datasets through its data lock and import bindings. Update those through its existing pinning command using the reviewed full HF commit. README links and moving branches are navigation, not execution inputs. Matching tags alone do not establish comparability; task/reference hashes, scorer identity, execution conditions and saved evidence must also match.