Where to Find Delphi Checkpoints on Hugging Face: A Complete Guide

Delphi checkpoints are hosted on Hugging Face under the marin-community/delphi collection, with specific model identifiers defined in the Marin repository's SFT configuration files.

The Marin repository (marin-community/marin) implements the Delphi scaling suite for language model research. All pre-trained model artifacts are published to a dedicated Hugging Face collection, while the training codebase maintains exact checkpoint references for reproducibility.

Locating the Official Delphi Collection on Hugging Face

The central repository for all Delphi model checkpoints is the marin-community/delphi collection on Hugging Face:

https://huggingface.co/collections/marin-community/delphi-69f93cbd09845c03b070bae9

This collection aggregates every released checkpoint from the Delphi training runs. Each entry follows a systematic naming scheme that encodes the compute budget and hyperparameters used during training.

Finding Specific Checkpoint Identifiers in the Marin Repository

While the collection hosts the artifacts, the Marin source code contains the precise identifiers required to reproduce experiments.

SFT Configuration Files

In experiments/sft/configs/delphi_1e22.py, the checkpoint for the 1e22 FLOP experiment is defined at lines 49-55:

DELPHI_1E22_PREPARED_CKPT = (
    "laion/delphi-1e22-p33m67-32p07b-lr0.67-54770ae7"
)

This constant references the exact Hugging Face repository ID (laion/delphi-1e22-p33m67-32p07b-lr0.67-54770ae7) used by the training scripts to load the base model for supervised fine-tuning.

Documentation in README.md

The repository's main documentation confirms the release location in README.md at lines 35-36, directing users to the Hugging Face collection for model downloads.

Downloading Delphi Checkpoints Programmatically

The Marin codebase leverages the huggingface_hub library to stream checkpoints directly from the Hub.

Complete Model Download

To download an entire checkpoint repository, including model weights, tokenizer, and configuration files:

from huggingface_hub import snapshot_download

# Checkpoint identifier from the collection

checkpoint_repo = "laion/delphi-1e22-p33m67-32p07b-lr0.67-54770ae7"

# Downloads all files to a local cache directory

local_path = snapshot_download(
    repo_id=checkpoint_repo, 
    revision="main"
)

print(f"Checkpoint available at: {local_path}")

Selective Tokenizer Download

If you only require the tokenizer files without downloading gigabytes of model weights:

from huggingface_hub import snapshot_download

tokenizer_path = snapshot_download(
    repo_id=checkpoint_repo,
    revision="main",
    allow_patterns=[
        "tokenizer*", 
        "special_tokens_map.json", 
        "added_tokens.json"
    ]
)

print(f"Tokenizer stored at: {tokenizer_path}")

Checkpoint Preparation Logic

The prepare_checkpoint_step function in experiments/sft/prepare_checkpoint.py (lines 45-48) demonstrates the production pattern used throughout Marin. This utility streams the raw checkpoint from the Hugging Face Hub and rewrites specific tensors required for the SFT training pipeline, rather than downloading the full artifact to local storage unnecessarily.

Summary

• Delphi checkpoints are published under the marin-community/delphi Hugging Face collection
• Specific identifiers like laion/delphi-1e22-p33m67-32p07b-lr0.67-54770ae7 are hardcoded in experiments/sft/configs/delphi_1e22.py
• Programmatic access uses snapshot_download from the huggingface_hub library
• Selective downloads can target only tokenizer files using the allow_patterns parameter
• Production workflows in prepare_checkpoint.py stream checkpoints directly from the Hub

Frequently Asked Questions

What is the exact URL for the Delphi Hugging Face collection?

The official collection is located at https://huggingface.co/collections/marin-community/delphi-69f93cbd09845c03b070bae9. This URL is referenced in the Marin repository's README.md at lines 35-36.

How do I download a specific Delphi checkpoint using Python?

Use the snapshot_download function from the huggingface_hub package. Pass the full repository ID (e.g., laion/delphi-1e22-p33m67-32p07b-lr0.67-54770ae7) as the repo_id parameter to download the complete model artifact to your local cache.

Where in the Marin codebase are checkpoint identifiers defined?

Checkpoint identifiers are defined as string constants in the SFT configuration files, specifically in experiments/sft/configs/delphi_1e22.py at lines 49-55. The constant DELPHI_1E22_PREPARED_CKPT stores the Hugging Face repository ID for the 1e22 FLOP experiment.

What is the naming convention for Delphi checkpoints?

Delphi checkpoints follow the pattern laion/delphi-{flops}-{params}-{tokens}-{lr}-{hash}, where flops indicates the compute budget (e.g., 1e22), params and tokens show the model scale and data mix ratios, lr specifies the learning rate, and hash provides a unique commit identifier.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}