YBenchmark Arena
mock-first · providers off

Benchmark Registry

Immutable, content-hashed benchmark versions. Each task records its scoring method, difficulty, contamination provenance, and tags.

Yuu Golden Regression Suite

golden

The permanent regression suite run against every Yuu release. Deterministic-first coverage across the capability taxonomy. Immutable at this version.

yuu-golden@1.0.0
hash c0889f0670d86668
24 tasks
scoring: numeric-tolerance ×7exact-match ×7unit-test ×3sql-result ×1regex ×5json-schema ×1
provenance: private ×24
TaskCategoryScoringJudgeDifficultyTags
g-reason-01Reasoningnumeric-tolerancedeterministicmediumarithmetic-trap, word-problem
g-reason-02Reasoningexact-matchdeterministiceasysyllogism
g-math-01Mathnumeric-tolerancedeterministiceasyorder-of-operations
g-math-02Mathnumeric-tolerancedeterministicmediumrates
g-code-01Codingunit-testdeterministiceasyarray, implementation
g-code-02Codingunit-testdeterministicmediumstack, parsing
g-debug-01Debuggingunit-testdeterministicmediumoff-by-one
g-sql-01SQLsql-resultdeterministicmediumgroup-by, aggregate
g-tool-01Tool Usenumeric-tolerancedeterministiceasysingle-tool
g-tool-02Tool Usenumeric-tolerancedeterministicmediumtool-selection
g-longctx-01Long Contextexact-matchdeterministicmediumneedle, single-hop
g-longctx-02Long Contextexact-matchdeterministicmediumneedle, single-hop, shallow
g-longctx-03Long Contextexact-matchdeterministicmediummulti-hop
g-longctx-04Long Contextexact-matchdeterministicmediumcontradiction, recency
g-longctx-05Long Contextexact-matchdeterministicmediumtemporal-order
g-instr-01Instruction Followingregexdeterministiceasyformat-constraint
g-instr-02Instruction Followingexact-matchdeterministiceasyexact-format
g-struct-01Structured Outputjson-schemadeterministicmediumjson, strict-format
g-halluc-01Hallucinationregexdeterministicmediumunanswerable, refuse-to-fabricate
g-plan-01Planningregexdeterministiceasyordering, dependencies
g-agentic-01Agenticnumeric-tolerancedeterministichardmulti-step, tool-chain
g-safety-01Safetyregexdeterministicmediumharmful-request, must-refuse
g-safety-02Safetyregexdeterministiceasybenign, over-refusal-check
g-data-01Data Analysisnumeric-tolerancedeterministiceasydescriptive-stats

Coding Depth

Algorithmic and implementation tasks graded by unit tests in the sandbox.

coding-depth@1.0.0
hash 3766b7e00cec3321
3 tasks
scoring: unit-test ×3
provenance: synthetic ×3
TaskCategoryScoringJudgeDifficultyTags
code-fizzbuzzCodingunit-testdeterministiceasyclassic
code-dedupe-stableCodingunit-testdeterministiceasyarray
code-lruCodingunit-testdeterministicharddata-structure, cache

SQL Analytics

Result-set-graded SQL over in-memory fixtures.

sql-analytics@1.0.0
hash bb4728240950565d
3 tasks
scoring: sql-result ×3
provenance: synthetic ×3
TaskCategoryScoringJudgeDifficultyTags
sql-region-revenueSQLsql-resultdeterministiceasygroup-by
sql-havingSQLsql-resultdeterministicmediumhaving, aggregate-filter
sql-avg-revenueSQLsql-resultdeterministicmediumaggregate, round

Structured Output

Schema adherence and semantic correctness of JSON responses.

structured-output@1.0.0
hash 69f5cb87549c8a17
3 tasks
scoring: json-schema ×3
provenance: synthetic ×3
TaskCategoryScoringJudgeDifficultyTags
struct-contactStructured Outputjson-schemadeterministicmediumextraction, nested
struct-enumStructured Outputjson-schemadeterministiceasyenum, classification
struct-nested-toolStructured Outputjson-schemadeterministichardnested-json, tool-result-synthesis

Long Context

Retrieval, multi-hop, temporal ordering, and summarisation over generated documents.

long-context@1.0.0
hash 5853f676ee880f22
3 tasks
scoring: exact-match ×2numeric-tolerance ×1
provenance: synthetic ×3
TaskCategoryScoringJudgeDifficultyTags
lc-temporalLong Contextexact-matchdeterministichardtemporal-order
lc-summary-fidelityLong Contextnumeric-tolerancedeterministicmediumsummarization, fidelity
lc-deep-needleLong Contextexact-matchdeterministicexpertneedle, deep, 24k

Research (LLM-judged)

Open-ended research answers graded by an LLM judge panel. Scores here are model-graded and reported separately from deterministic scores.

research-judged@1.0.0
hash ccc8645788047634
2 tasks
scoring: llm-judge ×2
provenance: synthetic ×2
TaskCategoryScoringJudgeDifficultyTags
research-tradeoffsResearchllm-judgemodel-gradedmediumarchitecture, open-ended
research-explainResearchllm-judgemodel-gradedeasyexplanation

Benchmark creator

New benchmarks are authored as immutable data with the defineSuite builder (see src/lib/benchmarks), or imported as JSONL/CSV. AI-generated candidate tasks are supported but must carry generated: true provenance so a generated task is never mistaken for a hand-verified one. Publishing a change to an existing benchmark creates a new version rather than mutating tasks in place — the content hash enforces this.