Oculus - Complete Training Repository

This repository contains the complete Oculus vision-language model including all training code, checkpoints, and documentation.

Quick Links

Model	Description	Link
Oculus-0.1-Instruct	Instruction-tuned for VQA/captioning	HuggingFace
Oculus-0.1-Reasoning	Chain-of-thought reasoning	HuggingFace
oceanir	Python SDK	PyPI

Installation

pip install oceanir

from oceanir import Oculus

model = Oculus.from_pretrained("OceanirAI/Oculus-0.1-Instruct")
answer = model.ask("image.jpg", "What is this?")

Architecture

Oculus combines state-of-the-art vision encoders with a powerful language model:

Vision Encoders

• DINOv3 ViT-H/16+ (facebook/dinov3-vith16plus-pretrain-lvd1689m)

  • Self-supervised vision transformer trained on LVD-1689M
  • 1024 hidden, 24 layers, 16 heads

• SigLIP2 (google/siglip2-base-patch16-224)

  • Vision-language contrastive model
  • 1152 hidden, 27 layers, 16 heads

Language Model

• LiquidAI LFM 2.5 1.2B Instruct (LiquidAI/LFM2.5-1.2B-Instruct-MLX-bf16)
  • 1.2B parameters, 1536 embedding dim
  • 131K vocab, 32K context window

Architecture Specs

Component	Specification
DINOv3	ViT-H/16+, 1024D, 24L, 16H
SigLIP2	Base, 1152D, 27L, 16H
Fusion	Concatenation → 2176D
Projector	2176 → 4352 → 1536
LFM 2.5	1.2B params, 1536D, 16L, 24H
Detection	80 classes (COCO)
Segmentation	150 classes (ADE20K)

Repository Structure

OceanirAI/Oculus/
├── config.json                    # Main model config
├── README.md                      # This file
│
├── oculus_unified_model/          # Model implementation
│   ├── __init__.py
│   ├── modeling_oculus.py         # OculusForConditionalGeneration
│   ├── configuration_oculus.py    # OculusConfig
│   └── processing_oculus.py       # OculusProcessor
│
├── training/                      # Training scripts
│   ├── train_oculus.py            # Base projector training
│   ├── train_detection.py         # Detection head training
│   ├── train_detection_extended.py
│   ├── train_instruction_tuning.py # Instruct variant
│   ├── train_reasoning_v2.py      # Reasoning variant
│   └── train_oculus_coco.py       # COCO training
│
├── logs/                          # Training logs
│   ├── training_instruct_v1.log
│   ├── training_reasoning_v2.log
│   └── training_v2_final.log
│
├── checkpoints/                   # Model checkpoints
│   ├── oculus/final/              # Base projector
│   │   ├── projector.npz          # Vision projector weights (~822MB)
│   │   └── config.json
│   │
│   ├── oculus_detection/final/    # Detection checkpoint
│   │   ├── projector.npz          # Projector weights (~800MB)
│   │   ├── heads.pth              # Detection heads (~35MB)
│   │   └── benchmark_results.json
│   │
│   ├── oculus_instruct_v1/        # Instruction-tuned VQA
│   │   └── vqa_model/
│   │       ├── model.safetensors  # BLIP VQA weights (~1.5GB)
│   │       ├── tokenizer.json
│   │       └── config.json
│   │
│   └── oculus_reasoning_v2/       # Reasoning VQA
│       └── vqa_model/
│           ├── model.safetensors  # BLIP VQA weights (~1.5GB)
│           ├── tokenizer.json
│           └── config.json
│
├── docs/                          # Documentation
│   ├── ARCHITECTURE.md
│   ├── BENCHMARK_README.md
│   └── TRAINING_ROADMAP.md
│
├── oculus_inference.py            # Inference script
├── demo_oculus.py                 # Demo script
├── benchmark_vlm.py               # Benchmarking
└── eval_benchmarks.py             # Evaluation

Training

Base Projector Training

python training/train_oculus.py

Detection Head Training

python training/train_detection.py

Instruction Tuning

python training/train_instruction_tuning.py

Reasoning Training

python training/train_reasoning_v2.py

Features

• Visual Question Answering (VQA) - Answer questions about images
• Image Captioning - Generate natural descriptions
• Object Detection - Detect with bounding boxes (80 COCO classes)
• Object Counting - Count objects via point prediction
• Semantic Segmentation - Pixel-level understanding (150 ADE20K classes)
• Chain-of-Thought Reasoning - Step-by-step thinking traces

License

Oceanir Research License v1.0

Permitted:

• Academic research
• Educational use
• Publishing papers with results
• Personal experimentation

Not Permitted:

• Commercial use
• Training commercial models
• Commercial products/services

For commercial licensing: licensing@oceanir.ai

Citation

@software{oculus2026,
  title={Oculus Vision-Language Model},
  author={OceanirAI},
  year={2026},
  url={https://huggingface.co/OceanirAI/Oculus}
}

Links

• Oculus-0.1-Instruct
• Oculus-0.1-Reasoning
• Oceanir SDK (PyPI)
• GitHub