NVIDIA just dropped NVEagle 🦅

Super impressive vision language model that comes in 7B, 13B and 13B fine-tuned on chat 💬
Model repositories: merve/nveagle-66d0705108582d73bb235c26
Try it: NVEagle/Eagle-X5-13B-Chat 💬 (works very well! 🤯)

This model essentially explores having different experts (MoE) for image encoder part of vision language model.
How? 🧐
The authors concatenate the vision encoder output tokens together, and they apply "pre-alignment" essentially fine-tune experts with frozen text encoder.

Then they freeze both experts and the decoder and just train the projection layer, and finally, they unfreeze everything for supervised fine-tuning ✨

In the paper, they explore different fusion strategies and vision encoders, extending basic CLIP encoder, and figure out simply concatenating visual tokens works well.
Rest of the architecture is quite similar to LLaVA. (see below the architecture)

reacted to singhsidhukuldeep's post with 👍 5 months ago

Post

2605

Good folks at Meta has just unveiled Llama 3.2, pushing the boundaries of language models and computer vision.

Even more interesting is how they trained this cutting-edge model:

1️⃣ Architecture:
Llama 3.2 uses an optimized transformer architecture with auto-regressive capabilities. The largest models (11B and 90B) now support multimodal inputs, integrating both text and images.

2️⃣ Training Pipeline:
• Started with pretrained Llama 3.1 text models
• Added image adapters and encoders
• Pretrained on large-scale noisy (image, text) pair data
• Fine-tuned on high-quality in-domain and knowledge-enhanced (image, text) pairs

3️⃣ Vision Integration:
• Trained adapter weights to integrate a pre-trained image encoder
• Used cross-attention layers to feed image representations into the language model
• Preserved text-only capabilities by not updating language model parameters during adapter training

4️⃣ Post-Training Alignment:
• Multiple rounds of supervised fine-tuning (SFT)
• Rejection sampling (RS)
• Direct preference optimization (DPO)
• Synthetic data generation using Llama 3.1 for Q&A augmentation
• Reward model ranking for high-quality fine-tuning data

5️⃣ Lightweight Models:
• Used pruning and distillation techniques for 1B and 3B models
• Structured pruning from Llama 3.1 8B model
• Knowledge distillation using Llama 3.1 8B and 70B as teachers

6️⃣ Context Length:
All models support an impressive 128K token context length.

7️⃣ Safety Measures:
Incorporated safety mitigation data to balance helpfulness and safety.

The result? A suite of models ranging from edge-friendly 1B parameters to powerful 90B parameter versions, capable of sophisticated reasoning across text and images. Llama 3.2 is set to revolutionize AI applications from mobile devices to enterprise-scale solutions.

What are your thoughts on these advancements? How do you see Llama 3.2 impacting your industry? Let's discuss in the comments!