From c088055827ef8d40d3b60103b19b21d5f80c522f Mon Sep 17 00:00:00 2001 From: Cemil ILIK Date: Sun, 18 Oct 2026 09:58:02 +0300 Subject: [PATCH 1/2] docs(usermanuals): say what the medical-qa-tr template really enables The templates reference (English and Turkish) said medical-qa-tr "ships with risk_assessment + safety eval pre-wired". The template's config has no `evaluation` and no `risk_assessment` block; only the sample answers carry a safety disclaimer. A user starting from a template in a regulated domain would have believed a gate was on. The row now says what is and is not enabled and what to add before any use beyond a demo. Refs #97 Co-Authored-By: Claude Sonnet 5.5 --- CHANGELOG.md | 2 ++ docs/usermanuals/en/reference/yaml-templates.md | 2 +- docs/usermanuals/tr/reference/yaml-templates.md | 2 +- 3 files changed, 4 insertions(+), 2 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 0d44dd02..ce87c3f0 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -68,6 +68,8 @@ All notable changes to ForgeLM are documented here. ### Fixed +- Correct the `medical-qa-tr` row of the YAML templates reference (English and Turkish): the template enables no safety + evaluation and no `risk_assessment`, so it no longer says it ships with both pre-wired (#97). - Fix the KTO bullet on the Alignment overview user-manual page (English and Turkish): it named the fields `(prompt, response, good?)`, but the KTO loader reads `prompt`, `completion` and a boolean `label` (#965). diff --git a/docs/usermanuals/en/reference/yaml-templates.md b/docs/usermanuals/en/reference/yaml-templates.md index 5fb6481a..f6fd2fd5 100644 --- a/docs/usermanuals/en/reference/yaml-templates.md +++ b/docs/usermanuals/en/reference/yaml-templates.md @@ -23,7 +23,7 @@ The five templates and the user journey each one optimises for: | `customer-support` | SFT | Multi-turn helpful + safe assistant — the most common production pattern. | | `code-assistant` | SFT | Code-completion / pair-programming model. | | `domain-expert` | SFT | BYOD (Bring Your Own Dataset) on a domain-specific corpus. | -| `medical-qa-tr` | SFT | Turkish-language medical QA (regulated domain — ships with `risk_assessment` + safety eval pre-wired). | +| `medical-qa-tr` | SFT | Turkish-language medical QA (regulated domain — the sample answers carry a safety disclaimer, but the template enables no safety evaluation and no `risk_assessment`; add an `evaluation.safety` block and a `risk_assessment:` block before any use beyond a demo). | | `grpo-math` | GRPO | Reward-model fine-tuning on math benchmarks (built-in format/length reward shaping). | The canonical template content is **the source of truth** — it is what diff --git a/docs/usermanuals/tr/reference/yaml-templates.md b/docs/usermanuals/tr/reference/yaml-templates.md index 9288a614..09719813 100644 --- a/docs/usermanuals/tr/reference/yaml-templates.md +++ b/docs/usermanuals/tr/reference/yaml-templates.md @@ -23,7 +23,7 @@ Beş şablon ve her birinin optimize ettiği kullanıcı yolu: | `customer-support` | SFT | Multi-turn yardımcı + güvenli asistan — en yaygın production pattern. | | `code-assistant` | SFT | Kod-tamamlama / pair-programming modeli. | | `domain-expert` | SFT | Domain-spesifik corpus üzerinde BYOD (Bring Your Own Dataset). | -| `medical-qa-tr` | SFT | Türkçe tıbbi soru-cevap (regüle alan — `risk_assessment` + safety eval ön-bağlı ship olur). | +| `medical-qa-tr` | SFT | Türkçe tıbbi soru-cevap (regüle alan — örnek cevaplar bir güvenlik uyarısı taşır, ama şablon güvenlik değerlendirmesini ve `risk_assessment`'ı etkinleştirmez; demo dışındaki bir kullanımdan önce bir `evaluation.safety` bloğu ve bir `risk_assessment:` bloğu ekle). | | `grpo-math` | GRPO | Matematik benchmark'ları üzerinde reward-model fine-tuning (yerleşik format/length reward shaping). | Kanonik şablon içeriği **doğruluk kaynağıdır** — `forgelm quickstart`'ın From bee88feb83f089a1e86c23b5780d0ecfe8a35895 Mon Sep 17 00:00:00 2001 From: Cemil ILIK Date: Sun, 18 Oct 2026 11:00:07 +0300 Subject: [PATCH 2/2] docs(usermanuals): do not imply the medical template becomes fit for clinical use Review findings on the medical-qa-tr row: adding the two blocks does not make the template suitable for clinical decision support, and an evaluation.safety block without enabled: true leaves the check off. The row now says the template is not for clinical decision support, to set evaluation.safety.enabled: true, and to fill in the risk_assessment block accurately. English and Turkish. Co-Authored-By: Claude Sonnet 5.5 --- docs/usermanuals/en/reference/yaml-templates.md | 2 +- docs/usermanuals/tr/reference/yaml-templates.md | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/docs/usermanuals/en/reference/yaml-templates.md b/docs/usermanuals/en/reference/yaml-templates.md index f6fd2fd5..75341a39 100644 --- a/docs/usermanuals/en/reference/yaml-templates.md +++ b/docs/usermanuals/en/reference/yaml-templates.md @@ -23,7 +23,7 @@ The five templates and the user journey each one optimises for: | `customer-support` | SFT | Multi-turn helpful + safe assistant — the most common production pattern. | | `code-assistant` | SFT | Code-completion / pair-programming model. | | `domain-expert` | SFT | BYOD (Bring Your Own Dataset) on a domain-specific corpus. | -| `medical-qa-tr` | SFT | Turkish-language medical QA (regulated domain — the sample answers carry a safety disclaimer, but the template enables no safety evaluation and no `risk_assessment`; add an `evaluation.safety` block and a `risk_assessment:` block before any use beyond a demo). | +| `medical-qa-tr` | SFT | Turkish-language medical QA (regulated domain — the sample answers carry a safety disclaimer, but the template enables no safety evaluation and has no `risk_assessment` metadata. It is not for clinical decision support, and adding these does not change that; before any use beyond a demo, set `evaluation.safety.enabled: true` and fill in the top-level `risk_assessment:` block accurately). | | `grpo-math` | GRPO | Reward-model fine-tuning on math benchmarks (built-in format/length reward shaping). | The canonical template content is **the source of truth** — it is what diff --git a/docs/usermanuals/tr/reference/yaml-templates.md b/docs/usermanuals/tr/reference/yaml-templates.md index 09719813..5baf2e0d 100644 --- a/docs/usermanuals/tr/reference/yaml-templates.md +++ b/docs/usermanuals/tr/reference/yaml-templates.md @@ -23,7 +23,7 @@ Beş şablon ve her birinin optimize ettiği kullanıcı yolu: | `customer-support` | SFT | Multi-turn yardımcı + güvenli asistan — en yaygın production pattern. | | `code-assistant` | SFT | Kod-tamamlama / pair-programming modeli. | | `domain-expert` | SFT | Domain-spesifik corpus üzerinde BYOD (Bring Your Own Dataset). | -| `medical-qa-tr` | SFT | Türkçe tıbbi soru-cevap (regüle alan — örnek cevaplar bir güvenlik uyarısı taşır, ama şablon güvenlik değerlendirmesini ve `risk_assessment`'ı etkinleştirmez; demo dışındaki bir kullanımdan önce bir `evaluation.safety` bloğu ve bir `risk_assessment:` bloğu ekle). | +| `medical-qa-tr` | SFT | Türkçe tıbbi soru-cevap (regüle alan — örnek cevaplar bir güvenlik uyarısı taşır, ama şablon güvenlik değerlendirmesini etkinleştirmez ve `risk_assessment` metadata'sı içermez. Klinik karar desteği için değildir ve bunları eklemek bunu değiştirmez; demo dışındaki bir kullanımdan önce `evaluation.safety.enabled: true` yap ve en üstteki `risk_assessment:` bloğunu doğru biçimde doldur). | | `grpo-math` | GRPO | Matematik benchmark'ları üzerinde reward-model fine-tuning (yerleşik format/length reward shaping). | Kanonik şablon içeriği **doğruluk kaynağıdır** — `forgelm quickstart`'ın