GDPval — file-producing deliverables¶
GDPval is a set of 220 long-horizon, file-oriented professional tasks. dockyard_rl has two paths, sharing one judge.
Text path¶
environments/gdpval_environment.py — single-turn, no sandbox. The model produces
a textual deliverable; it is scored against the task’s rubric_json by an LLM
judge (GDPvalRubricReward). The episode ends after this one scoring step.
Rubric grading is judgmental with no deterministic fallback — a judge
endpoint must be configured, or every sample is an execution_error. This is the
reduced path, intended primarily as an eval/validation signal.
Agentic path¶
environments/gdpval_agentic_environment.py — multi-turn, sandbox-backed, the
faithful path. Built on MultiTurnSessionEnvironment, the agent drives a
long-lived container and uses shell/python tools to produce real deliverable
files (xlsx, docx, pptx, pdf, …) under /workspace/deliverable, emitting one
```bash block per turn (or TASK_COMPLETE to finish).
On completion or budget exhaustion, the produced files are extracted to text
in-container (a manifest plus per-file content) and graded with the same
GDPvalRubricReward as the text path — so file- and format-oriented rubric
criteria that the text path structurally cannot satisfy become demonstrable to
the judge.
Configs: examples/configs/grpo_gdpval.yaml (text) and grpo_gdpval_agentic.yaml
(agentic).