Turn graded traces into curated training data.
What it does
Converts graded traces into curated SFT examples and DPO preference pairs for training.
When to use it
Use when graded traces or failure examples need conversion into training data, for rejection sampling, or for DPO pairs from passing and failing runs.
How to use it
Give it graded trace files and evaluation goldens; it writes curated JSONL training rows in SFT or DPO format.
What you provide
- Your files
What you need · 3
The prescribed goldens-holdout check is Python code and must run before a converted batch merges into the training set.
Requires traces already graded upstream by this skill; traces missing a verdict or reward must return there before conversion.
Requires this skill's Format Selection table as the source of exact output field names and its dataset-card structure for provenance.
Access · 0
No additional actions listed.