Rewrite synthetic data gen docs for Kiln Pro batch planning

[sfierro]

Updates the Synthetic Data Generation docs for the new Kiln Pro batch planning flow, and corrects drift in the existing (manual) content. Behavior was audited against the Kiln app source rather than inferred.

Accuracy fixes

Verified against the Kiln repo:

• Fine-tuning split tag is fine_tune_data, not fine_tuning_data
• Eval split tag names come from the eval's own tag filters; only the 80/20 ratio is fixed
• Eval templates were wrong. The docs claimed two options (Issue, Requirements) that you pick from. In reality there are four eval-derived templates (Requirements Eval, Desired Behaviour Eval, Appropriate Tool Use, Issue Eval — legacy), and Kiln auto-selects the one matching your eval's type
• Documented the six built-in templates (Fine Tuning, Toxicity, Bias, Maliciousness, Jailbreak, Factual Correctness)
• Refreshed model recommendations, which named models that are now out of date
• Noted the Data Guide applies at the topic and input stages only, not to output generation

Structure

The previous revision inserted the Pro flow alongside the untouched manual content, which left the two flows reading as unrelated and duplicated their shared steps.

• Both flows are now framed as the same four stages, with only the first differing (Kiln Pro plans the batch; manual builds a topic tree). Inputs, outputs, and saving are documented once instead of twice
• Templates and custom guidance are scoped to the manual flow. The template dropdown and per-stage guidance are manual-only in the app; Kiln Pro uses a single prefilled plan-level Guidance box. The old shared section implied they worked the same way
• Data gen model choice moved ahead of the build flows
• Clarified that Manual refers to how coverage is decided, not to hand-authoring data (matching the app's "Manual vs AI Planned" coverage framing)

Link fixes

• evaluations.md and fine-tuning-guide.md repointed to the current section
• Two of those links pointed at a #human-guidance anchor that did not exist and were already broken before this change
• All internal and external anchors into this page verified to resolve

Notes for review

• A few <!-- TODO IMAGE --> / <!-- TODO VIDEO --> placeholders mark screenshots still needed for the Pro flow
• The "Resolving Bugs & Issues" recipe was softened from "create an eval using the Issue template" to a template-neutral phrasing, since the Issue template is now legacy-only. Worth a look if there's a current canonical path there

🤖 Generated with Claude Code