Managing Hugging Face Models

Managing Hugging Face Models#

Treat models exactly like large datasets:

  • Pin your versions: Models update; always link a specific revision/commit.

  • Control your cache: Point downloads to fast local storage, not network drives.

  • Separate adapters: Save tiny LoRA weights, not entire base models.

Referring to Models

Never just write: “We used Llama-3.”

  • Use the exact repository ID: meta-llama/Meta-Llama-3-8B

  • Always specify the revision (commit SHA or tag).

Downloading and Storing Snapshots#

Control download location:

from huggingface_hub import snapshot_download

# Download exactly what you need to a specific folder
path = snapshot_download(
    repo_id="mistralai/Mistral-7B-v0.1",
    revision="26bca36b...",
    local_dir="data/raw/models/mistral"
)
Storing Adaptations (Fine-Tuning)

Do not duplicate base models!

  • Fine-tuning a 15GB model creates another 15GB model.

  • With PEFT / LoRA save only the adapter weights (~50MB).

  • Track these small adapters with standard Git or DVC.

Cluster Tip:

Set the HF_HOME environment variable to point to $SCRATCH or $TMPDIR to avoid crashing shared network drives!

Just be aware that the data will be purged!

Content