README.md

---
license: apache-2.0
pipeline_tag: zero-shot-image-classification
tags:
- datology
- clip
- vision
- OpenCLIP
- datacomp
- zero-shot-classification
---

# DatologyAI CLIP Classification Optimized ViT-B/32

**DatologyAI CLIP** is a state-of-the-art contrastive vision-language model that achieves superior performance through advanced data curation alone, without any architectural or training modifications. This classification-optimized ViT-B/32 model outperforms SigLIP2, MetaCLIP, and DFN on zero-shot classification benchmarks.

## Model Description

DatologyAI's CLIP model demonstrates that careful data curation can drive state-of-the-art performance without modifications to model architecture or training paradigms. Key achievements include:

- **76.91% ImageNet1k accuracy** (vs 74.0% for SigLIP2)
- **8x training efficiency** compared to standard approaches
- Trained on 13B curated image-text pairs from DataComp
- Standard CLIP architecture and training procedure

## Intended Uses

You can use this model for zero-shot image classification or as a vision encoder for VLMs and other vision tasks.

### Zero-shot Image Classification

```python
import torch
from PIL import Image
import open_clip

# Load model and preprocessing
model, _, preprocess = open_clip.create_model_and_transforms('hf-hub:DatologyAI/cls-opt-vit-b-32')
tokenizer = open_clip.get_tokenizer('hf-hub:DatologyAI/cls-opt-vit-b-32')

# Load image
image = preprocess(Image.open("path/to/image.jpg")).unsqueeze(0)

# Define candidate labels
labels = ["a dog", "a cat", "a bird"]
text = tokenizer(labels)

# Run inference
with torch.no_grad():
    image_features = model.encode_image(image)
    text_features = model.encode_text(text)
    
    # Normalize features
    image_features /= image_features.norm(dim=-1, keepdim=True)
    text_features /= text_features.norm(dim=-1, keepdim=True)
    
    # Calculate similarity
    similarity = (100.0 * image_features @ text_features.T).softmax(dim=-1)
    
# Get predictions
values, indices = similarity[0].topk(3)
for value, index in zip(values, indices):
    print(f"{labels[index]}: {value.item():.2%}")
```

### Image Encoding

```python
import torch
from PIL import Image
import open_clip

# Load model
model, _, preprocess = open_clip.create_model_and_transforms('hf-hub:DatologyAI/cls-opt-vit-b-32')
model.eval()

# Process image
image = preprocess(Image.open("path/to/image.jpg")).unsqueeze(0)

# Extract features
with torch.no_grad():
    image_features = model.encode_image(image)
    
print(f"Feature shape: {image_features.shape}")  # [1, 512]
```

## Training Procedure

DatologyAI's training pipeline focuses on sophisticated data curation techniques including:

1. **Improved target distribution matching** - Task-specific alignment of image features for classification
2. **Enhanced synthetic data generation** - Optimized caption generation for classification tasks
3. **Predictive metrics for curation quality** - Rapid iteration without full model training

The model uses standard CLIP training objectives with no architectural modifications.

## Training Data

The model was trained on 13B image-text (multi-epoch) curated from the **DataComp-XL** dataset using DatologyAI's proprietary curation pipeline. The curation process selected high-quality, classification-relevant subsets from the 10B available pairs in DataComp-XL.

## Evaluation Results

### Zero-shot Classification Performance

| Benchmark | DatologyAI | SigLIP2 | MetaCLIP |
|-----------|------------|---------|----------|
| ImageNet1k | **76.91%** | 74.0% | 67.7% |
| ImageNetv2 | **70.2%** | 67.1% | 60.4% |

### Training Efficiency
- Matches SigLIP2 performance with only **5B samples** (87.5% compute reduction)
- Matches MetaCLIP performance with only **1B samples** (92% compute reduction)

Full details see [blog post]().

## Model Details

- **Developed by:** DatologyAI
- **Model type:** CLIP (Contrastive Language-Image Pre-training)
- **Architecture:** Vision Transformer B/32
- **License:** Apache 2.0
- **Training framework:** OpenCLIP 2.24.0

## Technical Specifications

### Model Architecture
- **Vision Encoder:** ViT-B/32 (86M parameters)
  - Patch size: 32×32
  - Image size: 224×224
  - Embedding dimension: 512
- **Text Encoder:** 12-layer Transformer
  - Context length: 77 tokens
  - Vocabulary size: 49,408 (BPE tokenizer)

### Training Configuration
- **Optimizer:** AdamW (β1=0.9, β2=0.98, ε=1e-6)
- **Learning rate:** 5.0e-04 with cosine schedule
- **Weight decay:** 0.1
- **Batch size:** 32,768
- **Training samples:** 13B image-text pairs
- **Hardware:** Distributed training on H100 GPUs

## Citation

If you use this model, please cite:

```bibtex
@article{datologyai2025clip,
  title={CLIP Gets a Data Upgrade: Outperforming SoTA with Improved Data Curation Only},
  author={DatologyAI Team},
  journal={DatologyAI Blog},
  year={2025},
  url={https://datologyai.com/blog/clip-data-upgrade}
}
```

## Additional Information

For more details on our data curation methodology and comprehensive benchmark results, please visit our [blog post](https://datologyai.com/blog/clip-data-upgrade).

**Contact:** [team@datologyai.com](mailto:team@datologyai.com)

## Model Card Contact

DatologyAI Team - [team@datologyai.com](mailto:team@datologyai.com)