docs/usermanuals/en/concepts/data-formats.md: broken example — instructions format for SFT = {"prompt", "completion"} rows (DOC-D3B05-005) #232

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

Summary

The headline SFT format in the quick-reference table fails at training start after the model has loaded.

Primary finding

DOC-D3B05-005 — docs/usermanuals/en/concepts/data-formats.md: broken example — instructions format for SFT = {"prompt", "completion"} rows

severity High

Where. docs/usermanuals/en/concepts/data-formats.md:14,20-27,149 (Turkish mirror: docs/usermanuals/tr/concepts/data-formats.md:14,147) · type: broken example

The document says. instructions format for SFT = {"prompt", "completion"} rows

What the code does. The SFT path accepts only a text column, a messages column, or User/instruction + Assistant/output/response columns (data.py:483-488, 434-449). A prompt/completion file raises KeyError in prepare_dataset: "Dataset must contain 'User'/'instruction' and 'Assistant'/'output' columns … Detected format: prompt-only format. Suggested trainer: grpo" (reproduced with the tiny-qwen tokenizer). --dry-run does not catch it.

Code references. forgelm/data.py:434-449, forgelm/data.py:477-490, forgelm/data.py:40-41

User impact. The headline SFT format in the quick-reference table fails at training start after the model has loaded.

Proposed correction. Document the real single-turn SFT shapes, {"instruction": …, "output": …} (or response) and {"User": …, "Assistant": …, "System"?: …}, plus the pre-formatted {"text": …}, or extend data.py to map prompt/completion.

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.}