Home

An LLM reads your training curves and designs your next experiment.

pip install "neuropt[llm]"

Beats Optuna within 15 evals. Same budget, same search space (14 params), same CNN training on FashionMNIST. neuropt reads per-epoch training curves and reasons about why configs work — not just which ones scored well. See benchmarks for details.

Two ways to use it

Option 1: Define a search spaceOption 2: Give it a model

search_space = {
    "lr": (1e-4, 1e-1),                    # auto-detects log-scale
    "hidden_dim": (32, 512),                # auto-detects integer
    "activation": ["relu", "gelu", "silu"], # categorical
}

def train_fn(config):
    model = build_my_model(config["hidden_dim"], config["activation"])
    # ... train, return per-epoch losses for smarter LLM decisions ...
    return {"score": val_loss, "train_losses": [...], "val_losses": [...]}

model = torchvision.models.resnet18(weights="DEFAULT")

def train_fn(config):
    m = config["model"].to("cuda")  # deep copy with modifications applied
    # ... train ...
    return {"score": val_loss, "train_losses": [...], "val_losses": [...]}

neuropt walks the module tree — activations, dropout, norms, pooling — and builds a search space automatically. For pretrained models, it adds fine-tuning strategies too.

Then run:

neuropt run train.py                        # runs until Ctrl+C
neuropt run train.py --backend claude -n 50  # stop after 50

Or in a notebook:

from neuropt import ArchSearch

search = ArchSearch(train_fn=train_fn, search_space=search_space, backend="claude")
search.run(max_evals=50)

Reads training curves

The LLM sees full per-epoch train/val loss, not just a final score. It spots overfitting, underfitting, and LR issues from curve shapes.

Auto-introspects models

Give it a model and neuropt handles the rest — all the introspection and fine-tuning logic happens under the hood. You don't need to know or configure any of it.

Return anything

Your train_fn can return any metrics. Scalars become table columns, lists become per-epoch curves. The LLM sees everything.

Crash-safe & resumable

Everything is logged to JSONL. Kill it, restart it, it picks up where it left off. Works overnight, works in notebooks.

Works with

PyTorch models, XGBoost, LightGBM, Random Forest, any sklearn estimator. Supports Claude, OpenAI, and local Qwen backends.