alignment-overview describes KTO data as (prompt, response, good?), but the KTO loader reads prompt, completion and label #965

Summary

The "Alignment overview" user-manual page describes KTO as learning from "binary (prompt, response, good?) signals". Those are not the names the KTO loader reads: it requires prompt, completion and a boolean label. A reader who follows the page's wording and names their answer column response gets the same KeyError that #885 / #938 removed from the "Choosing a Trainer" table.

What happens

docs/usermanuals/en/concepts/alignment-overview.md:56 and the Turkish mirror at the same line (tr/concepts/alignment-overview.md:56, (prompt, response, iyi mi?)) name the KTO fields as response and good?.

_validate_trainer_columns raises KeyError: KTO trainer requires 'completion' and 'label' (boolean) columns … for any other shape (forgelm/data.py:600-607), and format detection routes only completion + label to KTO (forgelm/data.py:36-37). The KTO training page, the data-formats page and docs/guides/alignment.md already say completion + label; this bullet is the one place that still disagrees.

This is prose, not a column name in a table, which is why it was left out of #938 (one concern per PR).

Where

• docs/usermanuals/en/concepts/alignment-overview.md:56
• docs/usermanuals/tr/concepts/alignment-overview.md:56
• forgelm/data.py:600-607 (KTO column check, for reference)

Recommended fix

Name the real fields and keep the plain-language meaning of the label, in both languages. For example (English):

KTO — learns from binary (prompt, completion, label) rows, where label is true for a desirable response and false for an undesirable one. Use when you only have thumbs-up/down feedback.

The Turkish bullet keeps its structure and only swaps the field names (response → completion, iyi mi? → label, with the same true / false explanation). No wording beyond those names needs native review.

Out of scope

• The other items on the same page (DPO/SimPO memory multipliers, the sequence diagram) are tracked in [Wave 4] Low/info — Docs: docs/usermanuals/en/concepts/ #784. If a PR for [Wave 4] Low/info — Docs: docs/usermanuals/en/concepts/ #784 and this one touch the same lines (:54-56), merge one before the other.
• The {prompt, completion} → SFT row of the "Choosing a Trainer" table is a separate defect tracked in Documented SFT prompt/completion format is rejected with a KeyError that suggests trainer_type grpo (CR-A03-005) #403.

Acceptance criteria

• The KTO bullet names the columns the loader reads (prompt, completion, label) in the English page and its Turkish mirror, in the same PR.
• A ### Fixed entry is added under [Unreleased] in CHANGELOG.md.
• tools/check_bilingual_parity.py --strict, tools/check_bilingual_code_blocks.py --strict, tools/check_anchor_resolution.py --strict, tools/check_tr_links_prefer_mirror.py --strict and tools/check_usermanual_self_contained.py --strict pass.

Related

• Choosing-a-trainer table documents KTO rows as {prompt, response, label}, but the KTO loader requires a completion column (F-W1-DOCS-TRAIN-02) #885 / docs: fix KTO data-shape row to use completion column #938 — the same wording defect in the "Choosing a Trainer" table (fixed).
• [Wave 4] Low/info — Docs: docs/usermanuals/en/concepts/ #784 — other items on this page.

_{Found while reviewing PR #938. Verified against development at ea82d68.}