[Wave 3] Medium — TR mirror: docs/usermanuals/en/training/

Wave 3 · 2 finding(s) in 2 fix package(s). Mirror drift across small document pairs in this directory.

Findings

DOC-TR13-001 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/training/dpo.md: - Enable reference_free: true (less stable but skips the second model).
severity Medium · already tracked: KNOWN-OPEN C31-DOC-SCHEMA-PROSE docs/roadmap/phase-16-trust-surface-hardening.md:785-786
Where. EN dpo.md:128 ('Enable reference_free: true (less stable but skips the second model)') · TR dpo.md:116 ('reference_free: true.')
English text. - Enable reference_free: true (less stable but skips the second model).
Turkish text. - reference_free: true.
Which is correct. Both wrong: training has extra=forbid and no reference_free field (config.py:443-460), and both documents state two paragraphs earlier that ForgeLM does not expose it (EN dpo:114, TR dpo-tr:105). TR additionally drops the 'less stable' caveat and the verb, so it reads as a bare instruction.
Proposed correction (Turkish text for the -tr page). - reference_free ForgeLM'de yoktur; referans belleğini azaltmak için LoRA/QLoRA kullanın veya SimPO'ya geçin.
DOC-TR13-002 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/training/galore.md: So GaLore at r=256 (illustrative; the shipped default is r=128) lets you full-fine-tune a 7B model on a single 24 GB GPU — roughly the same…
severity Medium
Where. EN galore.md:65 ('…roughly the same VRAM as plain LoRA at full precision.') · TR galore.md:65 ('…kabaca full-precision LoRA ile aynı VRAM, ama her ağırlığa erişimle.')
English text. So GaLore at r=256 (illustrative; the shipped default is r=128) lets you full-fine-tune a 7B model on a single 24 GB GPU — roughly the same VRAM as plain LoRA at full precision.
Turkish text. … 7B modeli tek 24 GB GPU'da full fine-tune edebilirsiniz — kabaca full-precision LoRA ile aynı VRAM, ama her ağırlığa erişimle.
Which is correct. Neither: PEFT is applied unconditionally (model.py:609) so only LoRA adapters are trainable; TR adds an extra false assertion ('her ağırlığa erişimle').
Proposed correction (Turkish text for the -tr page). Yani GaLore, optimizer-state belleğini düşürür; ForgeLM'de PEFT adapter'ları yine eğitilir, bu nedenle tablodaki '%100 eğitilebilir parametre' ve '24 GB'a full fine-tune' iddiaları ölçülmemiş / geçersizdir.

Done when

Every finding above is fixed or closed with a reason (tick it here).
Code fixes carry a test; doc fixes update the Turkish mirror in the same PR.

_{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.}

Where to start

Thanks for taking a look. Please confirm the issue is unassigned and comment before starting so work does not overlap.

Start with these files on development:

Remove the unsupported reference_free instruction in both DPO pages. Correct the GaLore claim so it does not promise full-parameter training or measured 24 GB performance: ForgeLM still applies PEFT, and the schema allows GaLore alongside LoRA. Remove directly contradictory full-training claims within that manual pair; change no schema or trainer behavior.

Coordinate with #880: the DPO manual pair is shared. This is a docs correction, not an implementation of full fine-tuning.

Run the focused tests:

pytest tests/test_config.py tests/test_model.py --no-cov

Run these guards/checks as applicable to the changed files:

ruff format --check forgelm/config.py forgelm/model.py tests/test_config.py tests/test_model.py
ruff check forgelm/config.py forgelm/model.py tests/test_config.py tests/test_model.py
python3 tools/check_bilingual_parity.py --strict
python3 tools/check_bilingual_code_blocks.py --strict
python3 tools/check_anchor_resolution.py --strict
python3 tools/check_tr_links_prefer_mirror.py --strict
python3 tools/check_usermanual_self_contained.py --strict
python3 tools/check_usermanual_schema_drift.py --strict

If user-facing documentation changes, update the existing Turkish mirror in the same PR; keep commands, code and terminal output in English and structurally aligned. English-only contributor standards need no new mirror.

Open the PR against development with Fixes #642 once every finding is resolved or explicitly closed with a reason; use Refs #642 for partial work and list the resolved finding IDs.

I'm removing good first issue here and marking this help wanted, for two reasons. First, it needs natural Turkish written or reviewed by a native or near-native speaker: docs/standards/localization.md does not accept unreviewed machine translation (see its "No machine-translation output" rule). Second, this finding is tracked by Phase 16 step S15 (#901), which lands as one root-cause family with its own review rounds (see "Roadmap steps" in CONTRIBUTING.md). Please don't start on it without coordinating on #901 first.

Navigation Menu

Repository navigation

[Wave 3] Medium — TR mirror: docs/usermanuals/en/training/ #642

Findings

Done when

Where to start

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[Wave 3] Medium — TR mirror: docs/usermanuals/en/training/ #642

Description

Findings

Done when

Activity

cemililik commented on Oct 11, 2026

Where to start

cemililik commented on Oct 15, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions