This is the dataset of trained neural network checkpoints used to meta-train the NiNo model from https://github.com/SamsungSAILMontreal/nino/.

It contains 1000 models in total:

300 small convnets with 3 layers and 16, 32 and 32 channels (14,378 parameters in each model), trained on FashionMNIST (FM-16)
300 small convnets with 3 layers and 16, 32 and 32 channels (14,666 parameters in each model), trained on CIFAR10 (C10-16)
200 small GPT2-based transformers with 3 layers, 24 hidden units and 3 heads (1,252,464 parameters in each model), trained on LM1B (LM1B-3-24)
200 small GPT2-based transformers with 2 layers, 32 hidden units and 2 heads (1,666,464 parameters in each model), trained on LM1B (LM1B-2-32)

Each model contains multiple checkpoints:

2688 checkpoints per each model in FM-16 (corresponding to every 4 steps of Adam)
2513 checkpoints per each model in C10-16 (corresponding to every 4 steps of Adam)
124 checkpoints per each model in LM1B-3-24 (corresponding to every 200 steps of Adam)
124 checkpoints per each model in LM1B-2-32 (corresponding to every 200 steps of Adam)

In total, there are 1,609,900 model checkpoints.

The dataset also contains the training loss for each checkpoint, for FM-16 and C10-16 it also contains training accuracy, test loss, test accuracy.

The dataset corresponds to the first 4 columns (in-distribution tasks) in Table 1 below.

This Table is from the Accelerating Training with Neuron Interaction and Nowcasting Networks paper, see https://arxiv.org/abs/2409.04434 for details.

Example

UPDATE: To train the NiNo model more efficiently we also uploaded the dataset in the mmap format (in the mmap branch), which is much faster to read for large arrays. See https://github.com/SamsungSAILMontreal/nino/blob/main/dataset/dataset.py#L123 for the dataset code and https://github.com/SamsungSAILMontreal/nino/blob/main/README.md#training-nino for the training script examples.

Download and access trajectories of the fm-16 task (other three datasets: 'c10-16', 'lm1b-3-24', 'lm1b-2-32'). The original trajectories was saved as float16, but when loading huggingface may convert it to float32 or float64 depending on the version.

from datasets import load_dataset
import numpy as np

# Load the 'fm-16' task
dataset = load_dataset('SamsungSAILMontreal/nino_metatrain', 'fm-16')  # you can use streaming=True for checking out individual samples

# Once downloaded, access the t-th state in the i-th trajectory: 
print(np.array(dataset[str(i)][t]['data']).shape)  # (14378,)