docs: add prompt sensitivity devnote

[dhruvnathawani]

Summary

Add a dev note documenting the prompt sensitivity SDG pipeline used to generate diverse prompt variations for Nemotron training data across both SFT and RL.

What's in the post

1. Motivation: Why prompt sensitivity matters for model robustness (up to 15 percentage point accuracy swings from phrasing changes alone)
2. Prompt anatomy diagram showing the three variable components: preamble, problem (fixed), format instruction
3. Goal: reduce LLM sensitivity to prompt phrasing by generating diverse preambles and format instructions while keeping the core problem unchanged
4. Pipeline walkthrough: Seed preambles x format templates (cross-product) -> diversity samplers -> LLM preamble generation -> format instruction paraphrasing -> user prompt composition with placement ordering -> 4 quality judges -> YAML-driven training mixture integration
5. ASCII pipeline diagram showing the 5-stage flow (seed examples -> samplers -> LLM generation -> dual judges -> training mixtures)
6. Regex-paired format templates: 25+ answer formats (boxed, brackets, XML tags, asterisks, arrows, etc.), each paired with an extraction regex enabling both SFT diversity and RL reward parsing from a single pipeline
7. YAML-driven mixture config with majority_percentage control (25% canonical / 75% diverse)
8. Collapsible full source script using the DD config API (pip install + run)
9. Key takeaways on sampler-driven diversity, format compliance gating, and unified SFT/RL design

Files changed

1. docs/devnotes/posts/prompt-sensitivity.md (updated from draft to full PR)
2. docs/devnotes/posts/prompt_anatomy.png (new)

[greptile-apps]

Greptile Summary

Added comprehensive devnote documenting the prompt sensitivity pipeline used to generate diverse prompt variations for Nemotron training. The document explains how Data Designer's sampler-driven approach creates 1,000+ validated preambles spanning multiple diversity dimensions (tone, verbosity, format, etc.) to improve model robustness against prompt phrasing variations. Includes working code examples, ASCII diagrams, YAML configuration samples, and details on regex-paired format templates that enable both SFT and RL training from a single pipeline.

Confidence Score: 5/5

• Safe to merge - documentation-only changes with no code modifications
• Documentation is well-written, technically accurate, properly formatted, includes working code examples following project conventions, and successfully integrates into the existing docs structure
• No files require special attention

Important Files Changed

Filename	Overview
docs/devnotes/posts/prompt-sensitivity.md	Comprehensive devnote documenting prompt sensitivity SDG pipeline with clear explanations, working code examples, ASCII diagrams, and practical integration details
docs/devnotes/posts/prompt_anatomy.png	Clean visual diagram showing prompt anatomy with color-coded sections (preamble, problem, format instruction, output)
mkdocs.yml	Single line addition correctly integrating the new devnote into the documentation navigation

_{Last reviewed commit: bfca665}