Skip to content

Latest commit

 

History

History
49 lines (36 loc) · 2.18 KB

File metadata and controls

49 lines (36 loc) · 2.18 KB

Training data, and what is kept

Two stages at the far end of the line: what a scored run can be turned into, and what is worth keeping afterwards.

pairs — preference data from scored runs

masterwork pairs RUN_DIR --axis grounding --out pairs.jsonl \
  --min-gap 0.25 --gap-from band.json

On-policy preference pairs, cut from cells the benchmark already scored. The agent produced both sides; nothing is written for it.

--min-gap is required and has no default. A pair whose two sides differ by less than the judge's own spread pairs the judge's noise rather than the agent's behaviour, and training on that teaches the noise. Measure the spread first — same input, sampling seed varying, n≥3 — and set the gap above it. --gap-from records which measurement justified the number, so a set can be traced back to it.

Two refusals, both deliberate:

  • No pairs from a self-judged run. The benchmark stamps those, and a stamp that travels with the claim is the whole point. --allow-self-judged exists, is recorded on every pair it produces, and should be rare.
  • No pairs below the gap. There is no flag for this one.

--sft cuts winners only, as a supervised set, with --floor for the minimum score to count as a winner.

--strip-system drops the identity from the prompt. The difference matters: with the system text in place you teach behaviour conditional on that prompt; without it you teach the behaviour itself. Which one you want depends on whether the identity will be present at inference.

retain — what a run leaves behind

masterwork retain check RUN_DIR
masterwork retain archive RUN_DIR --to archive/

check reports the size of a run directory and never deletes anything — deciding what to remove is not the line's call. archive copies out the kilobyte-sized evidence: the run record and the report, the two files a disagreement six weeks later is settled with. Everything else under a run is re-derivable.

A run that grows past max_run_bytes (default 50 MB) is reported in its own record, not trimmed. A line that quietly deleted evidence to stay tidy would be deciding, on its own, which parts of the record stop mattering.