Skip to content

Gdpval Harness

GDPval empirical grounding harness for G6 job agents.

Cluster: Uncategorised | Type: component | MCP Tools: None

Overview

Empirical grounding harness that evaluates G6 job agents against the OpenAI GDPval benchmark dataset. Loads tasks from HuggingFace (openai/gdpval, 2.29 GB cached), maps occupations to job agent components, routes each task to the appropriate agent, scores deliverables against rubrics with optional LLM-as-judge, and emits JSONL run records. Includes calibration tools for computing per-job prior success probabilities from benchmark runs.

When to use:

  • Evaluating job agent quality against a standardised occupational benchmark
  • Computing calibrated priors for Bayesian confidence estimation
  • Regression testing after job agent modifications
  • Generating evidence for tier advancement decisions

Works well with: All job_* components, align_evals, self_training, adapt_bayesian

Public API

GdpvalHarnessInput(BaseModel)

Field Type Default
op str required
parameters dict[str, Any] Field(default_factory=dict)

GdpvalHarnessOutput(BaseModel)

Field Type Default
op str ''
result dict[str, Any] Field(default_factory=dict)
message str ''

GdpvalHarnessBlock(AIBlock)

AIBlock wrapper for GDPval empirical grounding harness tasks and scoring.

Methods:

infer(input: GdpvalHarnessInput) -> Result[GdpvalHarnessOutput]

GDPvalLoader

Constructor:

Parameter Type Default
cache_dir str \| Path \| None None
local_override str \| Path \| None None

Methods:

iter_tasks(limit: int | None = None) -> Iterator[dict]

Yield GDPval task dicts.