You can deploy this model and start earning money today!

fillo-rinaldi/
ViT-L-14-datacomp_xl_s13b_b90k

Name: ViT-L-14-datacomp_xl_s13b_b90k
Availability: InStock

open_clip

mit

Views

No views yet

Fine-Tuned OpenCLIP ViT-L-14 Checkpoints (DataComp-XL (datacomp_xl_s13b_b90k))

This repository contains full fine-tuned OpenCLIP ViT-L-14 checkpoints, each fine-tuned on a single downstream vision dataset starting from the DataComp-XL (datacomp_xl_s13b_b90k) pretrained weights. All models were trained by freezing the text encoder and fine-tuning only the visual backbone. These checkpoints are used as inputs for model merging and rebasin experiments in the Merge-and-Rebase project.

20 vision datasets, each with two checkpoints:

Dataset	Epochs
SUN397	14
Cars	35
RESISC45	15
EuroSAT	12
SVHN	4
GTSRB	11
MNIST	5
DTD	76
CIFAR100	6
STL10	6
Flowers102	147
OxfordIIITPet	82
PCAM	1
FER2013	10
EMNIST	2
CIFAR10	6
Food101	4
FashionMNIST	5
RenderedSST2	39
KMNIST	5

Each dataset folder contains:

full_best_ep.pt — checkpoint with the best validation accuracy
full_last_ep.pt — checkpoint from the final training epoch

Hyperparameters

All models were trained with the following shared configuration:

Hyperparameter	Value
Training strategy	Full fine-tuning (all parameters)
Fine-tuning scope	Visual backbone only (text encoder frozen)
Optimizer	AdamW
Learning rate	1e-5
Weight decay	0.1
Batch size	128
LR scheduler	Cosine (decay to 0)
Gradient clip norm	1.0
Early stopping	Disabled
Seed	42
Precision	fp32
Validation split	10% of training data

Backbone Details

Property	Value
Model	OpenCLIP `ViT-L-14`
Pretrained weights	`datacomp_xl_s13b_b90k`
Embedding dimension	512 (768)
Number of parameters	~428M

Source Code

These checkpoints were produced by merge_and_rebase/finetune/train_vision.py. The exact training configuration is in finetune/configs/vision_vitl14.yaml.

Usage

python

from huggingface_hub import hf_hub_download
import open_clip

checkpoint = hf_hub_download(
    repo_id="fillo-rinaldi/ViT-L-14-datacomp_xl_s13b_b90k",
    filename="SUN397/full_best_ep.pt",
    repo_type="model",
)

model, _, preprocess = open_clip.create_model_and_transforms(
    "ViT-L-14",
    pretrained="datacomp_xl_s13b_b90k",
    checkpoint_path=checkpoint,
)

fillo-rinaldi/ViT-L-14-datacomp_xl_s13b_b90k