docs/usermanuals/en/concepts/data-formats.md: broken example — For mixed formats, use separate files and reference both via data.dataset_name_or_path + data.extra_datasets (DOC-D3B05-009) #235

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

Summary

Following the advice crashes training. A multi-format corpus needs separate stages (pipeline:), not extra_datasets.

Primary finding

DOC-D3B05-009 — docs/usermanuals/en/concepts/data-formats.md: broken example — For mixed formats, use separate files and reference both via data.dataset_name_or_path + data.extra_datasets

severity High

Where. docs/usermanuals/en/concepts/data-formats.md:152-153 (Turkish mirror: docs/usermanuals/tr/concepts/data-formats.md:151) · type: broken example

The document says. For mixed formats, use separate files and reference both via data.dataset_name_or_path + data.extra_datasets

What the code does. _merge_extra_datasets concatenates the splits (data.py:512-546), producing union columns padded with None (DPO primary + SFT extra gave columns prompt/chosen/rejected/messages). The DPO run then crashed with AttributeError: 'NoneType' object has no attribute 'endswith', exit 2 (repro).

Code references. forgelm/data.py:512-546

User impact. Following the advice crashes training. A multi-format corpus needs separate stages (pipeline:), not extra_datasets.

Proposed correction. Say: "All files in one stage must share the trainer's format; to combine SFT and preference data use a pipeline: with separate stages."

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.}