Design Experiment

You help users plan experiments for fine-tuning and evaluating LLMs. Create a plan that specifies the complete workflow from training through evaluation, verifies resources, and documents all steps in a structured YAML configuration.

Your Task

Guide the user through designing their experiment by asking questions, verifying resources, and creating a comprehensive experiment_summary.yaml file that documents the complete plan.

Prerequisites Check

Quick existence check — claude.local.md must be present before designing an experiment, because every output path and SLURM default is read from it:

ls -la claude.local.md

If missing, stop and tell the user to run /ck-setup first — that skill walks them through claude.local.md.template interactively. Do not try to write a claude.local.md from this skill.

If claude.local.md is present but you suspect drift (placeholders not filled in, fields that look stale), suggest the user run /ck-setup in validate mode to get a structured health check rather than trying to validate inline here.

Dependency version check

Run before proceeding to catch stale envs (user pulled new pins but didn't re-run pip install -e .):

python scripts/check_env.py

• Exit 0: proceed.
• Exit 1: show the printed STALE ENV table to the user, ask whether to pip install -e . first or continue anyway.

Workflow

Follow the three-stage process:

1. Parameter Selection → param_selection.md

Guide the user through 9 interactive steps to gather all experiment parameters:

1. Determine experiment location
2. Understand the experiment (scientific question, variables, and whether tabular-to-text data generation is needed)
3. Confirm tool choices (torchtune for preparation, inspect-ai for evaluation)
4. Design training runs (models, datasets, hyperparameters)
5. Design evaluation runs (tasks, epochs, evaluation matrix)
6. Establish naming (experiment name, run names)
7. Verify resources (models, datasets, eval scripts exist)
8. Get approval (validate first, then present)
9. Create files (proceed to generation stage)

See param_selection.md for:

• Complete question flow for each step
• Auto-detection logic for experiment location
• Resource verification commands
• Conversation patterns

2. Validation → validation.md

Before presenting plan to user (step 8), validate completeness:

• ✓ All YAML sections present and properly structured
• ✓ All run names follow convention
• ✓ All parameters documented (variables and controls)
• ✓ Evaluation plan is consistent (0-indexed epochs, base vs fine-tuned)
• ✓ controls.system_prompt is set — the single source for training and eval (no separate eval copy to match)
• ✓ All resources verified (or noted as prerequisites)

See validation.md for:

• Complete validation checklist
• Common issues to check
• How to handle missing prerequisites

3. Experiment Generation → experiment_generation.md

After user approves, create output files:

1. experimentsummary.yaml - Structured experiment configuration (use templates/experimentsummary.yaml)
2. logs/design-experiment.log - Human-readable audit trail (see logging.md)

Then ask about next steps (scaffold-experiment?).

See experiment_generation.md for:

• File creation instructions
• YAML formatting guidance
• Next steps conversation pattern
• Prerequisites handling

Cross-Cutting Concerns

Logging → logging.md

IMPORTANT: Throughout paramselection and generation, create detailed log at {experimentdir}/logs/design-experiment.log.

What to log:

• ✓ Resource verification (ls, du, df commands and results)
• ✓ Prior run searches (if performed)
• ✓ Decisions (naming, recipe, configuration)
• ✓ File creation

Format: Plain text with timestamped action entries

See logging.md for:

• Complete log format specification
• All action types with examples
• Entry structure and timestamp format
• When to log during workflow

Templates → templates/

Reference materials for output generation:

• templates/experiment_summary.yaml - YAML schema and structure for experiment plan

Important Reminders

• BEFORE generating experimentsummary.yaml: You MUST read templates/experimentsummary.yaml first. Do not freestyle the YAML structure — use the template schema exactly.
• vislabel is required in evaluation matrix entries for visualization support (see paramselection.md Step 5)
• Dataset format terminology: Describe JSON datasets as "JSON with input/output keys" - never invent format type names
• Use paths from claude.local.md for models, datasets, scratch directories
• Always verify resources exist before finalizing plan (log all verification)
• System prompt has a single source - set controls.systemprompt; it propagates to both training and eval, so parity is automatic (per-task variation: evaluation.tasks[].systemprompt)
• Prompt can be swept - the user prompt defaults to controls.prompt but can vary per-task (evaluation.tasks[].prompt — eval-only prompt sweep) or per-run (runs[].parameters.prompt — a fine-tune trains on its own prompt). Author the structure, don't make the user hand-edit configs. See param_selection.md → Prompt Sweeps
• Epochs are 0-indexed - Use [0, 1, 2] in evaluation matrix
• Base models and eval-only checkpoints use epochs: null, fine-tuned models use epochs: [0, 1]
• controls.datasettype is required ("chatcompletion" | "text_completion") — read by torchtune at training time and propagated to drive chat-template choice at eval time for every run type, including eval-only
• Document tool choices in YAML - torchtune for training, inspect-ai for evaluation
• Handle missing resources gracefully - note as prerequisites, don't block the plan
• If inspect-ai task doesn't exist - note that create-inspect-task skill should be run first
• Generate YAML, not Markdown - Use structured YAML format with proper indentation

Module Organization

This skill uses the param_selection → validation → generation pattern:

Pattern: Three action verbs (selection, validation, generation) matching scaffold/run skills, plus cross-cutting logging and templates.

See README.md for: Complete pattern documentation and rationale.