Domains
Forms
Mechanism
Breaks down how the skill works and what it produces, so you can quickly judge whether it fits your scenario and value.
When graded evaluation traces or production logs exist (with verdicts/rewards from eval-harness-first) and need to become SFT examples or DPO preference pairs; when applying rejection sampling or building pairs from passing/failing runs.
eval/goldens.jsonl and runs/<run-id>/results.json with task_id, verdict, reward.JSONL training rows (SFT/DPO) shaped for dataset-curation input, with provenance and hygiene guarantees — closing the eval-to-training flywheel.
Use Cases
An ML team has 10k graded traces from eval-harness-first with pass/fail verdicts and reward scores. They use this skill to rank passing traces by reward, keep the top fraction as SFT rows, and build DPO pairs from same-task pass-vs-fail trajectories with rejected at μ−2σ. Secrets are scanned, goldens held out, and provenance attached. The cleaned JSONL feeds dataset-curation, saving weeks of manual labeling and improving model quality via high-signal data.
A production system logs failed model outputs that humans later correct. Using this skill, the team routes expert-corrected failures directly into the SFT set without reward thresholds, masks bad steps in multi-step traces, and applies judge-delta filtering to shrink pair volume. They dedup against existing training data and emit a dataset card with run_id/trace_id provenance. This turns costly incident fixes into reusable training signal, reducing repeat failures.
Skill Relationships
Dependency relationships read as "the upper tier points to the lower tier." The current Skill sits in the middle tier — above are Skills that depend on it, below are Skills it depends on.
Tier 1 · These Skills Use Me
Tier 2 · Current Skill
Tier 3 · I Use These Skills
Browse this skill's relationships within its skillset. Click a node to switch the side panel; use the search box to jump to any skill.
Skill File