| { |
| "$schema": "https://json-schema.org/draft/2020-12/schema", |
| "title": "MyPCBench Task Definition", |
| "description": "Schema for a MyPCBench evaluation task (the graded source: tasks/final/<bucket>/*.rubrics.json and the compiled tasks/final/all_tasks_with_grading.json). Each task tests whether a desktop agent can reason over the canonical persona's digital footprint. Grading is rubric-only (LLM-as-judge); programmatic checks were retired in an earlier revision. Note: tasks/final/all_tasks.json is a deliberately grading-free convenience list (no `grading` key) for runs that score offline via judge_results.py, so it does not validate against this schema's `required`.", |
| "type": "object", |
| "required": ["id", "category", "instruction", "grading"], |
| "properties": { |
| "id": { |
| "type": "string", |
| "pattern": "^[a-z_]+-[a-z0-9]+$", |
| "description": "Unique task identifier, e.g. retrieval-f001" |
| }, |
| "category": { |
| "type": "string", |
| "enum": [ |
| "retrieval", |
| "aggregation", |
| "contradiction", |
| "situated_action", |
| "preference_inference", |
| "counterfactual", |
| "long_horizon", |
| "cua_only", |
| "hard_app" |
| ], |
| "description": "Internal task category label" |
| }, |
| "instruction": { |
| "type": "string", |
| "description": "The natural-language task instruction given to the agent" |
| }, |
| "persona": { |
| "type": "string", |
| "enum": ["michael_scott"], |
| "description": "Optional. Present in the per-bucket source files; the canonical persona is supplied to the runner at run time via --persona (default michael_scott), so the flat eval files omit it." |
| }, |
| "grading": { |
| "type": "object", |
| "required": ["type", "rubrics"], |
| "properties": { |
| "type": { |
| "type": "string", |
| "enum": ["llm_judge"], |
| "description": "Rubric-only LLM-as-judge grading" |
| }, |
| "rubrics": { |
| "type": "array", |
| "items": { |
| "type": "object", |
| "anyOf": [ |
| { "required": ["criterion"] }, |
| { "required": ["requirement"] } |
| ], |
| "properties": { |
| "rubric_id": { "type": "string" }, |
| "criterion": { |
| "type": "string", |
| "description": "Natural-language criterion the judge evaluates over the full trajectory" |
| }, |
| "requirement": { |
| "type": "string", |
| "description": "Alias for criterion used in some buckets" |
| }, |
| "type": { "type": "string", "enum": ["llm_judge"], "default": "llm_judge" }, |
| "weight": { "type": "number", "minimum": 0, "maximum": 1 } |
| } |
| }, |
| "description": "LLM-as-judge rubric criteria" |
| } |
| } |
| }, |
| "time_limit_seconds": { |
| "type": "integer", |
| "description": "Soft time budget for the task in seconds" |
| }, |
| "horizon": { |
| "type": "string", |
| "description": "Interaction-horizon label (e.g. H1, H2)" |
| }, |
| "app": { |
| "type": "string", |
| "description": "Primary app the task targets (used by the smoke sets)" |
| }, |
| "apps_involved": { |
| "type": "array", |
| "items": { "type": "string" }, |
| "description": "Apps the task is expected to touch (analysis only, not enforced)" |
| }, |
| "difficulty": { |
| "type": "string", |
| "enum": ["easy", "medium", "hard"], |
| "description": "Estimated difficulty level" |
| } |
| } |
| } |
|
|
