nanovlm-COCO-VQAv2

This is a fine-tuned NanoVLM (Nano Vision-Language Model) trained on Mixed (COCO Captions + VQAv2) using Modal.com's cloud infrastructure.

Model Details

• Base Model: lusxvr/nanoVLM-222M
• Model Size: 222M parameters
• Architecture: Vision Transformer (SigLIP) + Small Language Model (SmolLM2)
• Training Platform: Modal.com (A100 GPU)
• Training Date: 2025-07-06

Architecture Components

• Vision Encoder: SigLIP-B/16-224 (85M parameters)
• Language Model: SmolLM2-135M
• Modality Projection: Pixel shuffle projection layer
• Total Parameters: ~222M

Training Details

Dataset

• Type: Mixed (COCO Captions + VQAv2)
• Description: A balanced combination of COCO image captions and VQAv2 question-answering pairs
• Size: 5,000 samples
• Multi-image Support: No

Training Configuration

• Batch Size: 8 (effective: 32)
• Training Steps: 500
• Learning Rate (MP): 0.00512
• Learning Rate (Backbones): 5e-05
• Model Compilation: Enabled
• Gradient Accumulation: 4 steps

Model Configuration

• Vision Model: google/siglip2-base-patch16-256
• Language Model: HuggingFaceTB/SmolLM2-360M-Instruct
• Image Size: 256x256
• Max Sequence Length: 1024
• Image Token Length: 64

Usage

Quick Start

from models.vision_language_model import VisionLanguageModel
from PIL import Image
import requests

# Load the model
model = VisionLanguageModel.from_pretrained("pgryko/nanovlm-COCO-VQAv2")

# Load an image
url = "https://huggingface.co/datasets/mishig/sample_images/resolve/main/tiger.jpg"
image = Image.open(requests.get(url, stream=True).raw)

# Generate a response
response = model.generate(
    image=image,
    prompt="What do you see in this image?",
    max_length=50
)
print(response)

Training Infrastructure

This model was trained using Modal.com's serverless GPU infrastructure:

• GPU: NVIDIA A100-40GB
• Training Time: ~60-75 minutes (including dataset preparation)
• Cost: ~$6-8 USD
• Platform: Modal.com serverless compute

Reproducibility

To reproduce this training:

# Using the integrated Modal approach
python modal/submit_modal_training.py \
  --build_dataset \
  --dataset_type mixed \
  --dataset_limit 5000 \
  --batch_size 8 \
  --max_training_steps 500 \
  --compile \
  --push_to_hub \
  --hub_model_id your-username/your-model-name

Monitoring

Training metrics and logs are available on Weights & Biases:

• Project: piotr-gryko-devalogic/nanovlm-modal

Limitations

• Context Length: Limited to 1024 tokens
• Image Resolution: Fixed at 256x256 pixels
• Language: Primarily English
• Domain: General vision-language tasks (performance may vary on specialized domains)

Ethical Considerations

This model inherits potential biases from its training datasets (COCO, VQAv2). Users should be aware of potential limitations in:

• Representation of diverse populations
• Cultural and geographic biases
• Object and scene recognition across different contexts

Citation

@misc{pgryko_nanovlm_COCO_VQAv2,
  title={NanoVLM Fine-tuned on Mixed (COCO Captions + VQAv2)},
  author={Modal.com Training Pipeline},
  year={2024},
  url={https://huggingface.co/pgryko/nanovlm-COCO-VQAv2}
}

Acknowledgments

• Base Model: nanoVLM by HuggingFace
• Training Platform: Modal.com for serverless GPU compute
• Datasets: Microsoft COCO and VQAv2 teams
• Infrastructure: NVIDIA A100 GPU via Modal.com

---

This model was trained using an automated pipeline on Modal.com. For questions or issues, please refer to the nanoVLM repository.