haipipe-task-for-fit

name	haipipe-task-for-fit
description	model-fitting task-folder specialist: scaffolds {NN}_<name>/ task-folders that fit/train a model (full hyperparam config, real GPU sweep, checkpoint to _WorkSpace/5-ModelInstanceStore/). NOT algorithm dev -- see /haipipe-task-for-algo. Called by /haipipe-task when task-type=fit. Cross-references /haipipe-nn-tuner and /haipipe-nn-instance.
argument-hint	[project_id] [group] [task-name]
allowed-tools	Bash, Read, Write, Edit, Grep, Glob, Skill
metadata	{"version":"0.1.2","last_updated":"2026-07-04","summary":"model-fitting task-folder build specialist."}

Skill: haipipe-task-for-fit

Scaffolds a model-training task-folder. Full training config, heavy outputs to _WorkSpace/5-ModelInstanceStore/, designed for cross-run comparison and paper-grade results.

Invocation modes: interactive (human steers; missing fields get ASKed) OR headless (haipipe-task-creator-agent calls this skill during Phase 2: Build, then authors the <TASK>.py body). Always end with the structured return block (status / task_folder / run_name / files).

Not the same as task-algo

See ../../2_nn/haipipe-task-for-algo/SKILL.md for the full comparison. Short version: algo-dev = smoke test, training = real run.

What this scaffolds

tasks/A{NN}_<group_name>/                    ← A-series group (model-run)
└── {NN}_<task_name>/
    ├── {NN}_<task_name>.py
    ├── configs/
    │   └── 5_model_<name>.yaml              seeded from ref/config-seed.yaml
    ├── runs/
    │   └── 5_model_<name>_<variant>.sh
    ├── results/
    │   └── <run>/                           model_path.txt + metrics.json (light)
    ├── notebooks/
    └── sbatch/                              optional, for GPU-partitioned sweep

Group letter default: A (model-run). Heavy outputs land in: _WorkSpace/5-ModelInstanceStore/.

Cross-reference to pipeline skill

/haipipe-nn-tuner defines the hyperparameter search space; /haipipe-nn-instance materializes a ModelInstance from a tuner sweep. This skill scaffolds the example task that drives both.

1. /haipipe-nn-algo — algorithm class exists.
2. /haipipe-nn-tuner — author the sweep.
3. /haipipe-task-for-fit — scaffold the example task.
4. Run sweep → ModelInstance → /haipipe-task-for-eval.

Scaffold flow

See fn/scaffold.md for the detailed step-by-step. Summary:

1. Identify project + task-group.
2. Collect metadata (NN, name, type-specific extras, _meta block).
3. Create skeleton (.py, configs/, runs/, results/, notebooks/).
4. Seed config from ref/config-seed.yaml.
5. Copy run-script from ../../haipipe-task/ref/run-sh-template.sh.
6. Suggest next via cross-skill link.
7. Emit return contract.

Return contract

status:    ok | blocked | failed
summary:   2-3 sentences on what was scaffolded
artifacts: [paths created]
next:      suggested next command (/haipipe-nn-tuner or run.sh)

Lessons learned

Canonical lessons live in LESSON.md (same folder): L1 temporal-leak filter, L2 PID/PatientID format, L4 empty prediction_results.json = silent bug, L5 -p SKIP_TRAINING "true" fast iteration, L6 what the Step-8 reproducibility check actually tests. Read it before building a fit task; do not restate its content here.

Quick pointers:

• ExampleFn builder lives in the same-numbered task as training (same number = same stage); builder templates at code/scripts/haibuilder/5-instance/ (workspace-dependent path).
• prediction_results.json must be non-empty after training (LESSON.md L4).

Workflow plan

When /haipipe-task plan targets an existing task-folder of this type, the generated plan-script YAML should follow the type-specific sample:

ref/workflow-plan-sample.yaml     ← script-level phases for this type
../../haipipe-task/ref/workflow-template.yaml  ← task-level template (Run/Gate1/Gate2)

Schema source of truth: task/haipipe-workflow/ref/plan-schema.md