Perplexity raises $73.6M & new small models: LLaVA-ϕ a 2.7B Phi-2 based multimodal model

Week 2 of Coding with Intelligence

📰 News

• Phi-2 from Microsoft becomes MIT licensed

• Perplexity raises $73.6M from IVP

  Does Google Search have some real competition on their hands? Try their RAG powered search and judge yourself: https://www.perplexity.ai/

📦 Repos

• Lilac: data curation tool for LLMs

• TACO: HumanEval alternative by BAAI

  GPT-4 apparently scores 31.5 pass@1 on easy and 2 pass@1 for very hard. By Shandong Normal University and Peking University.

• More details on 1.6T Switch Transformer from Google

  Updated HF page and updated source code: https://github.com/google-research/t5x

• AutoAWQ: implements the AWQ algorithm for 4-bit quantization with a 2x speedup during inference

📄 Papers

• Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models

  Bootstrapping better models!

• DeepSeek LLM paper: Scaling Open-Source Language Models with Longtermism

  Interesting section about tool-integrated reasoning performance on several benchmarks like GSM8K MATH MGSM-zh and CMath.

• Mixtral of Experts

  Paper from the Mistral team about their Mixtral 8x7B model

• LLaVA-ϕ: Efficient Multi-Modal Assistant with Small Language Model

  Llava based on Phi-2 - extra relevant with the recent MIT licensing of Phi-2 by Microsoft.

• Tencent trains Llama 2 Pro 8.3B: outperforms WizardCode/WizardMath

  Code & weights at https://github.com/TencentARC/LLaMA-Pro

• Steering Llama 2 via Contrastive Activation Addition

• Google AI releases Instruct-Imagen paper: instruction tuning for image generation

📱 Demos

• Auffusion: Leveraging the Power of Diffusion and Large Language Models for Text-to-Audio Generation

  Remarkably useful for content creators 🙌🏻

📚 Resources

• Chess-GPT's Internal World Model

  Really cool interpretability work, probing for the world model/decision model of an LLM trained on chess games

• Understanding Mixtral-8x7b: a visual exploration

• LLM Training and Inference with Intel Gaudi 2 AI Accelerators

  Good post by Databricks with lots of interesting tables & figures.

---

Want more? Follow me on Twitter! @ricklamers