Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →sylvanus4/automl-train
Schema-gated hyperparameter search (HPO/AutoML) skill for any training backend. Owns the closed loop — baseline eval, a validated search space, N trials, metric scrape, next-config recommendation, and an independent final eval of the best checkpoint — while deferring the single training run to whatever trainer you already have (Kubeflow Trainer v2, Ray Tune, SLURM, a local subprocess).
Passed 18 of 21 test cases with the skill, and 3 of 21 without it.
and 29 more in the full report
16 fixed3 passed either way3 still failing0 broke
Tokens, turns and time describe resource use, not a dollar price.
One of 16 test cases the skill fixed. Same model, same prompt, with and without the skill.
A trial failed in our HPO loop due to an S3 credential expiration during dataset download. What CLI tool parses trial logs and what standard tag is assigned?
| Checked | Without skill | With skill |
|---|---|---|
| Recommends executing failure_classify.py classify --log to parse the trial log. | Missed | Passed |
| Identifies image_cred or infra as the failure classification category. | Missed | Passed |
The instructions your model reads when the skill loads.
A thin orchestration layer that turns a single training run into a disciplined
hyperparameter search. The philosophy is borrowed from NVIDIA TAO's
tao-run-automl skill (Apache-2.0): gate before you spend GPU. What is new
here is that the four gates are owned by small deterministic scripts, so the
model recommends the next hyperparameters and nothing else — schema validation,
enum normalization, metric aggregation, and the promotion decision are code.
search_space_schema.py must pass — a valid, bounded space with no zero-training-step configs — or the loop cannot start.hpo_gate.py --check baseline requires a recorded reference metric.--check budget enforces max_concurrent * gpus_per_trial <= gpu_cap; launch_review.py estimates the wall-clock and refuses to start if it blows a stated limit.failure_classify.py tags each failed trial (data / image-cred / infra / spec-schema / model-code); if consecutive trials share a systemic cause, the loop stops and reports instead of exhausting the GPU budget.schema-gate -> baseline(+gate) -> launch-review(wall-time, +gate) ->
for each trial from trial_config_gen:
submit to YOUR trainer -> scrape metric (or classify failure) ->
if consecutive same-cause failures: stop early ->
select_best -> final-eval on best -> final-gate -> report -> promote
The model owns: writing the schema, choosing the strategy (random / ASHA / Hyperband), and reading gate results. Everything else is deterministic.
1/eta each rung with more resource.python scripts/search_space_schema.py --schema search-space.json # gate #1
test cases passed, vs 3 / 21 without
More skills from the same source and nearby in the registry, with their headline benchmark lift.