How the dataset is built¶

Each example is a hunk of code at one patch set, the reviewer comments anchored inside it, and the same hunk at the next patch set: what the developer wrote in answer to the review. The label is never guessed, since the next patch set is the answer. Most of what a review host holds is not such a pair, so most of it is removed, by rules fixed before any outcome was read.

The rules, in order¶

Collection. Only merged changes, created on or after 2024-10-01, after the base model's release. Each host is read at the pace its robots.txt or written terms set, by one client, and every account is replaced by a salted pseudonym as the data arrives.

Comments. A comment is dropped when it sits on the commit message or on the patch set as a whole (Gerrit's pseudo-files, not code anyone revised), when the change's own author wrote it (it can describe the edit it precedes), when a bot account wrote it (bots enforce written rules, not tacit ones), when its whole text is an acknowledgement such as Done or LGTM, when it has no line anchor, when no later patch set answered it, or when the next patch set changed nothing around its line.

Hunks. A hunk is dropped when there is nothing on one side of the edit, or when it grew or shrank past the size rules, since neither is a refinement a model can be asked to make.

Bots. A comment from an account the host tags as a bot is dropped when the comment is read. OpenStack's months were collected before that tag was kept at ingestion, so there the rule could not run; Wikimedia's build found no inline comment from a tagged account, since its bots post change-level messages. Some bots comment inline under an ordinary account, so a comment is also removed when its text matches a template read from the bot's own source (on Qt, its Sanity Bot, flake8 output and its security-review integration); applied to OpenStack's examples, it found none.

Examples. An example is dropped when its next patch set only rebased the change rather than reworking it, or when its target was written by the reviewer (Gerrit's suggested edit or its Fix applied reply, already in the prompt), or when only acknowledgements or a bot's template text are left once those comments are removed. An example whose next patch set's kind was never recorded is kept without the rebase check and counted.

Duplicates. Exact copies, MinHash near duplicates of the normalized before and after text, and shared boilerplate are removed across the whole organization, so the same pair is never in two windows, and a change found under two numbers keeps one.

Fetch errors. A change whose comments the host would not return, or a comment whose diff it would not return, is dropped and counted rather than stopping the build.

Windows. Examples are split by when the change was created: pilot (2024-10), train (2024-11 to 2025-08) and development (2025-09 to 2025-10). The test window (2025-11 to 2026-10) is defined and hashed now and fetched only after in-principle acceptance, so it is not counted here.

What each rule removed¶

Counts from each organization's frozen build (datasets/results/dataset-funnel.json), each row in its own unit. A comment removed at refinement may belong to an example then dropped for having nothing else left, so those two steps overlap. A counter a build did not record reads as not counted. Qt's column is its fetch under the 2026-10-15 grant, built skipping the changes whose listing records no comment once its comment-count probe held (research log, 2026-10-17).

What the rules cannot see¶

A sample of the kept examples was labeled by hand and by two model raters for whether the rewrite answers its comment. On OpenStack about four in five do; the rest stay in the corpus and are scored in every arm, so they dilute a contrast rather than favor one adapter (label-audit-v2-openstack.json, research log 2026-10-15). Edits a rebase brought into the next patch set also stay; the test window records Gerrit's rebase attribution so both hypotheses can be read again without them.