Zeeb — Video-LLM

Train from scratch with real data

1. Phase 1: Train VQ-VAE on 10K real images (COCO/imagenette)
2. Phase 2: Tokenize 10K image-text pairs through trained VQ-VAE
3. Phase 3: Fine-tune OLMo 2 1B + LoRA on 5K tokenized samples
4. Phase 4: Push trained model to EeshaAI/zeeb

Checkpoints saved to persistent storage (survives Space restarts). Training takes several hours on CPU.