Skip to content

triage 215 apply_fail records in sweep #64: dataset noise vs git-apply strictness #96

Description

@nikolay-e

Since 7711da2 the eval fns gate on apply_as_commit and record status=apply_fail instead of silently scoring a garbage HEAD~1..HEAD diff (previously these looked like "ok"). Sweep #64 interim data shows 215 such records.

Needed: sample ~20 of them from the run-64 checkpoints and classify —

  1. dataset noise (gold patch legitimately does not apply to base_commit, e.g. bad revision pin or pre-normalized patch) -> exclude those instance_ids upstream in benchmarks/manifests/ and note in SPLIT_REPORT;
  2. strictness (git apply --index rejects something a whitespace-tolerant apply would take) -> consider --whitespace=nowarn-class flags, NOT --3way (removed deliberately, see apply_as_commit docstring);
  3. mixed -> both.

Either way the denominator of every affected cell changes slightly; document the outcome in STATS_PLAN exclusions.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions