You can deploy this model and start earning money today!

yamboo/
Griffin_models

safetensors

graph-ml

yamboo/Griffin_datasets_joint_v65

yamboo/Griffin_datasets_single_pretrain_v3

2505.05568

apache-2.0

Views

No views yet

Griffin: Pretrained Checkpoints

This repository contains various pretrained checkpoints for the Griffin model. The paper is at Link

Checkpoints

The checkpoints are organized as follows:

bash

./checkpoints/
├── single-completion # Pretrained single table completion model.
├── single-sft # Pretrained single table SFT model. Used in main experiments.
└── transfer # Pretrained transfer model. Used in transfer experiments.
    ├── commerce-1 # Split name.
       ├── FULL # RDB-SFT setting name. This one used in main transfer experiments.
       ├── MIXED # RDB-SFT setting name. Used in ablation in RDB-SFT setting.
       └── LIMITED # RDB-SFT setting name. Used in ablation in RDB-SFT setting.
    ├── commerce-2 # Same as above.
       ├── FULL
       ├── MIXED
       └── LIMITED
    ├── others-1
       ├── FULL
       ├── MIXED
       └── LIMITED
    └── others-2
       ├── FULL
       ├── MIXED
       └── LIMITED

How to use

To get started, you will need to have the model's architecture defined in your code, provided in Github Repo. You can then use the huggingface_hub library to download a specific checkpoint and load its weights.

python

import json
import torch
from huggingface_hub import hf_hub_download
import accelerate

# Assume 'GriffinModel' is your model's class definition
# from your_project_position.hmodel import GriffinMod

# 1. Define the repository ID and the specific file you want to load
repo_id = "yamboo/Griffin_models"
# Example: Loading the main single-table SFT model
checkpoint_path = "single-sft/model.safetensors"
config_path = "single-sft/config.json"


# 2. Download the checkpoint file from the Hub
model_weights_path = hf_hub_download(repo_id=repo_id, filename=checkpoint_path)
model_config_path = hf_hub_download(repo_id=repo_id, filename=config_path)
config = json.load(open("config.json", "r"))

# 3. Instantiate your model and load the weights. We use accelerate to align with Github repo experiment pipeline.
model = GriffinMod(**config)  # Make sure to pass any required config
accelerate.load_checkpoint_in_model(model, model_weights_path)