Provider and attribution
LMArena Leaderboard Dataset
- Registry state
- registered
- Rights state
- approved
- Provider group
- lmarena
Contains data from the Arena Leaderboard Dataset by Arena, licensed under CC BY 4.0
Release planning · no model scores
Exact V3 collection, facet, metric, lineage, identity and rights mapping for Document arena from LMArena Leaderboard Dataset.
Provider and attribution
Contains data from the Arena Leaderboard Dataset by Arena, licensed under CC BY 4.0
V3 evidence records
A repeated upstream benchmark can map to several collections. Each record below preserves its own task role, facets, capability atoms, subject identity, metrics, lineages and exclusions.
supporting evidence · document-comprehension/lmarena/document-arena
Which models are best at document comprehension and extraction?
Collection facets
Field and entity extraction · objective · 25%
Extract requested values exactly and preserve missingness.
Tables and figures · objective · 25%
Reason over tabular and visual document content without losing units or labels.
Cross-document reconciliation · hybrid · 25%
Join facts across sources while retaining provenance and conflicts.
Requirements and document QA · hybrid · 25%
Answer precise questions and cite where the answer came from.
Capability atoms
Subject identity
Usable evidence lineages
lmarena-document-style-control/foundation-modelExact metric contract
| Metric | Lineage | Subject | Protocol | Use |
|---|---|---|---|---|
document_style_control.overallLMArena Document (style controlled) — Overall | lmarena-document-style-control | foundation-model | LMArena Document Arena style-controlled Bradley-Terry rating | usable |
Exclusions and blockers
No mapping-level exclusion is recorded; the page-level release gate still applies.
Mapping note
Preference does not replace assertion-level correctness checks.
supporting evidence · summarization-grounding/lmarena/document-arena
Which models are best at summarization and grounded knowledge work?
Collection facets
Meeting notes · hybrid · 25%
Capture decisions, owners, deadlines, and unresolved items.
Long-document summaries · hybrid · 25%
Preserve important coverage and caveats across long source material.
Grounded question answering · hybrid · 25%
Answer only from supplied evidence and expose insufficient support.
Factual consistency · objective · 25%
Avoid contradictions, invented details, and attribution drift.
Capability atoms
Subject identity
Usable evidence lineages
lmarena-document-style-control/foundation-modelExact metric contract
| Metric | Lineage | Subject | Protocol | Use |
|---|---|---|---|---|
document_style_control.overallLMArena Document (style controlled) — Overall | lmarena-document-style-control | foundation-model | LMArena Document Arena style-controlled Bradley-Terry rating | usable |
Exclusions and blockers
No mapping-level exclusion is recorded; the page-level release gate still applies.
Mapping note
Internal source-grounded checks remain primary.