docs/usermanuals/en/training/dpo.md: false claim — forgelm audit catches/flags chosen == rejected collisions and near-duplicates in preference data (DOC-D3B05-010) #236

Wave 1 · theme D1-10: Documentation-only fixes · highest severity High · 1 finding(s) · effort docs fix

Summary

Users skip their own check, and the 'loss converges near zero' pitfall goes undiagnosed.

Primary finding

DOC-D3B05-010 — docs/usermanuals/en/training/dpo.md: false claim — forgelm audit catches/flags chosen == rejected collisions and near-duplicates in preference data

severity High

Where. docs/usermanuals/en/training/dpo.md:103,145,208 (Turkish mirror: docs/usermanuals/tr/training/dpo.md:94) · type: false claim

The document says. forgelm audit catches/flags chosen == rejected collisions and near-duplicates in preference data

What the code does. No chosen/rejected check exists in forgelm/data_audit. For preference rows, near-dup and PII scanning use only prompt (_types.py:78). Repro: identical chosen/rejected rows were not flagged.

Code references. forgelm/data_audit/_types.py:78, forgelm/data_audit/_streaming.py:116-165

User impact. Users skip their own check, and the 'loss converges near zero' pitfall goes undiagnosed.

Proposed correction. Remove the claim and give a manual check (e.g. jq -c 'select(.chosen==.rejected)' data.jsonl).

Turkish mirror. the Turkish page has the same error; fix both in the same PR

Acceptance criteria

• Each statement is re-checked against the code at fix time.
• Documentation listed above is corrected in English and in its Turkish mirror in the same PR.
• Every finding in this issue is resolved or explicitly closed with a reason.

_{Found by the October 2026 ForgeLM code and documentation review against commit cab2947. Code links point at that commit; line numbers may have moved since. Finding IDs (CR-…, DOC-…, TECH-…) are stable references for this backlog.}