Shared Data/Model Repository

4.3. Shared Data/Model Repository#

We host several popular ML datasets and models on the Kempner AI cluster. This approach reduces the need for multiple transfers of the same data/model by researchers and provides a central, read-only repository for all Kempner Institute users to access for their ML workflows. Only the admin team has write access, but users can submit requests for popular data/models. After a careful review, we may place a copy in the shared data/model repository. The current path on the cluster is:

DATA_PATH=/n/holylfs06/LABS/kempner_shared/Lab/data
MODEL_PATH=/n/holylfs06/LABS/kempner_shared/Lab/model

Note

We will develop a web interface later for data and model discovery.

4.3.1. The current list of ML models#

  • CodeLlama

    • Path: $MODEL_PATH/models--codellama--CodeLlama-7b-hf (see on HuggingFace)

      • Size: 16 G

  • EleutherAI

    • Path: $MODEL_PATH/models--EleutherAI--pythia-160m-deduped (see on HuggingFace

      • Size: 435 M

    • Path: $MODEL_PATH/models--EleutherAI--pythia-70m-deduped (see on HuggingFace

      • Size: 195 M

  • OpenAI

    • Path: $MODEL_PATH/models--gpt2 (see on HuggingFace)

      • Size: 4.5 M

  • Google

    • Path: $MODEL_PATH/models--t5-base (see on HuggingFace)

      • Size: 3.4 M

4.3.2. The current list of ML datasets#

  • c4_original

    • Path: $DATA_PATH/c4_original

      • Subfolders:

        • preprocessed (434 M)

        • raw (157 G)

      • Description: The original version of the “Colossal Clean Crawled Corpus” (C4) dataset, designed for training natural language processing models.

  • dolma

    • Path: $DATA_PATH/dolma

      • Subfolders:

        • preprocessed (6.8 T)

        • raw (5.9 T)

      • Description: Dolma is a dataset of 3 trillion tokens from a diverse mix of web content, academic publications, code, books, and encyclopedic materials.

  • imagenet_winter21_whole

    • Path: $DATA_PATH/imagenet_winter21_whole

      • Subfolders:

        • winter21_whole.tar.gz (1.3 T)

      • Description: An updated version of the ImageNet dataset, containing a wide variety of annotated images for visual object recognition, collected during the winter of 2021.

Contents