looptimum

GPT training-parameter search

Ten evaluations tested four training controls. The selected model had lower held-out loss and fewer parameters than the fixed baseline.

Observed campaign results

10/10

Successful evaluations

4 initialization + 6 guided

1.02%

Lower held-out loss

Selected versus fixed baseline

~25%

Fewer parameters

Selected versus fixed baseline

Observed results

The selected model's held-out loss index was 98.98 against a fixed baseline of 100.00. Its parameter-count index was 75.4 against 100.0. This single-seed, small-budget campaign does not establish statistical significance or global optimality.

Search parameters

The search varied four controls within defined bounds. All ten evaluations completed successfully.

Learning rate Microbatch size Gradient accumulation Model depth

Evaluation method

Looptimum proposed the settings, and the client-owned training workflow returned a loss index measured on the fixed held-out data. The controller recorded the result before choosing the next trial.

Read the public evidence

Normalized campaign figures

Normalized held-out loss across ten evaluations in an anonymized GPT training campaign

Normalized objective progression

The guided phase produced the two strongest observed candidates; lower indexed loss is better.

Zero-based indexed comparisons of held-out loss and parameter count for a fixed baseline and selected candidate

Baseline versus selected candidate

The selected candidate combined a modest loss improvement with a separate reduction in model parameters.

Search progression

  • The guided phase found the two strongest observed candidates, with loss indices of 99.10 and 98.98.
  • Later trials explored the remaining bounds without improving that result.
  • This campaign does not establish a general advantage of guided over random search.

Limits of this result

  • Anonymized campaign with one seed and a small evaluation budget.
  • The observed result does not establish statistical significance or global optimality.
  • The search does not isolate the causal effect of any one control.
  • Transfer to another workflow, broader model quality, and production readiness were not assessed.
  • About 25% fewer parameters does not imply a matching reduction in runtime or cost.

Model-training pilot assessment

A project review covers the adjustable training parameters, evaluation metric, and resource cost per run.

Discuss a pilot