TIGER-Lab/WebsiteBenchArtifacts

WebsiteBench

70 tasks × 9 models. Start in benchmark/. Only the current evaluation protocol contributes scores. The versioned JSON files are the current catalog; Excel is no longer a runtime or update dependency.

benchmark/<task>/<model>/<context>/
  score.json                 # published score, metric and verification level
  trajectory.tar.zst         # ATIF + referenced images
  website.tar.gz             # generated website (.tar.zst for normalized exports)
  reference.tar.gz           # exact bound reference, when available
  evaluation.json.zst         # valid raw evaluation matching the published fraction
  provenance.json            # run identity, hashes, limitations and prior claims

An absent material has a trajectory.MISSING.json, website.MISSING.json, or evaluation.MISSING.json in the same folder. A run without a valid score has score.UNSCORED.json, including the reason. These markers contain no fabricated scores. Different contexts stay in separate folders; folder count is not the number of independent runs.

200K / 1M are reported context-window settings: approximately 200,000 / 1,000,000 tokens. They are separate run configurations, independent of the evaluation protocol. A larger context does not guarantee a higher score. A published context label does not by itself verify the runtime limit; configuration limitations remain in each run’s provenance.

The active cohort excludes Untappd, RoomSketcher, Workable and edX for the recorded task/evaluation problems. All four tasks and their materials are preserved under history/excluded-tasks/; they contribute to no current score, ranking, plot or coverage count. The exact 70-task list is sources/cohort.json.

What needs attention

• Missing trajectories — task/model pairs without a located baseline trace.
• Missing website code — task/model pairs without bound baseline code.
• No current valid score — includes closed or invalid evaluations; this does not automatically call for a rerun.
• Context-level missing materials and unscored contexts keep context restrictions visible.
• One-page coverage and all active pairs.

The default Overview missing-score list has 9 cells. Overview uses explicit per-model display contexts from display-contexts.json: Qwen uses its reported 200K run; GLM/DeepSeek retain their 1M browsing defaults. Eight explicitly listed GLM tasks use their available 200K publication, as requested; all other GLM selections use the 1M default. No maximum-score selection is used. Mixed columns and individual cells show actual contexts. The 1M-only list (36 cells) and 200K-only list (68 cells) are context-specific comparisons, not a model-wide missing-score claim. The model summary below counts a score in any recorded context.

Book1 Qwen reconciliation preserves the original 74-task audit: 71 published 200K fractions match the canonical score files; Apple, RoomSketcher and Untappd have no published valid fraction. The blank model-name cell for Apple is explicitly attributed to this single-model sheet.

Model	Traces	Websites	Published scores (any context)
Claude Opus 5.5	70/70	70/70	70/70
Claude Opus 5	70/70	69/70	69/70
GPT 5.6 Luna	70/70	70/70	70/70
Grok 4.6	70/70	70/70	70/70
GLM 5.3 Flash	70/70	70/70	70/70
Kimi K3	70/70	70/70	68/70
DeepSeek V4.1 Flash	70/70	70/70	67/70
Qwen 3.8 Max	70/70	68/70	69/70
GPT 6.1 Sol	70/70	69/70	68/70

Leaderboard

Table/Pareto data is recalculated from the exact same 70-task/context selection as Waffle. Overall and T1/T2/T3 are arithmetic means of the available per-task fractions; missing/invalid results are not zero. Every metric includes its coverage. Reported resource means preserve estimates/partial-usage qualifications. Pareto uses the same task subset for both coordinates. The auditable per-task inputs are in data/leaderboard-inputs.json.

Reading scores

score.json separates the reported fraction, T1/T2/T3, scoring protocol, exact recording ID, and evidence level. published_declaration means a current source reports the score; matching_valid_evaluation means a valid equal-case-v1 evaluation matches it; selected_run_and_cases_verified also checks explicit selection and case identities. Duplicate copies of the same evaluation are one piece of score evidence. Original 1000-case / functional-only scores and earlier publications without verified new-protocol evidence live in history/benchmark/, never in current score fields or Waffle. Their current folder is marked needs_current_evaluation. A missing score is never zero. All recorded raw evaluations, including invalid/old ones, remain separately labelled under evaluations/; evaluation.json.zst is specifically a valid new-protocol evaluation matching the current fraction.

The final migration reads all 1,126 Harbor tab records (959 rows within the original 74-task cohort), all 1,170 workbook notes (including all 934 old-Missing traces notes), and existing published claims. Repeated source rows are provenance, not extra runs. Explicit current invalid/closed results supersede stale scores. Empty placeholder rows do not erase existing published results. Original values, source cells and limitations are retained in provenance.json, migration provenance, and history/. No maximum-score selection is used. Rounded Sol context labels 258k/258.4k are one context; original labels remain recorded. Context configuration that was not verified is explicitly identified.

Website and updates

The website imports the small, checksummed site exports at a pinned HF revision. Update the relevant canonical score/material file, regenerate the derived tables/site exports, validate, and publish one catalog revision. There is no need to edit Excel. Large payloads reuse verified blobs; downloading the entire repository is unnecessary. Older source directories and objects/ remain available as archive paths.

python tools/export_unified_tables.py --root .
# Website checkout:
python tools/sync_hf_catalog.py --revision <40-character-HF-commit>
node tools/prerender.js

Source presence does not certify that a website launches unchanged. Known replay limitations and original licensing/attribution still apply. External credentials were redacted during normalization; hashes document the retained bytes. Reference exports are not borrowed across recordings.

Citation

See CITATION.bib and CITATION.cff.