instruct-eval VS Emu

I modified declare-lab's instruct-eval scripts, add support to VLLM, AutoGPTQ (and new autoGPTQ support exllama now), and test the mmlu result. I also add support to fastllm (which can accelerate ChatGLM2-6b.The code is here https://github.com/declare-lab/instruct-eval , I'd like to hear any errors in those code.

Well, I ran it with exactly the same parameters I ran Vicuna 7b, although I ran Vicuna with llama.cpp. while PJ can only be ran with ggml (I don't have a GPU). And Vicuna looped only when temperature reached 0. Given how hard it loops, I think it is some bug with ggml. Testers claim it should be close to 7b alpaca/vicuna:https://github.com/declare-lab/flan-eval

For a list of other instruction tuned models, you can check out this benchmark here: https://github.com/declare-lab/flan-eval

Great resource! I’ve recently also benchmarked many of the popular instruction models here: https://github.com/declare-lab/flan-eval

I really dislike this approach of announcing new models that some companies have taken, they don't mention evaluation results or performance of the model, but instead talk about how "transparent", "accessible" and "supportive" these models are.

Anyway, I have benchmarked stablelm-base-alpha-3b (the open-source version, not the fine-tuned one which is under a NC license) using the MMLU benchmark and the results are rather underwhelming compared to other open source models:

* stablelm-base-alpha-3b (3B params): 25.6% average accuracy

* flan-t5-xl (3B params): 49.3% average accuracy

* flan-t5-small (80M params): 29.4% average accuracy

MMLU is just one benchmark, but based on the blog post, I don't think it will yield much better results in others. I'll leave links to the MMLU results of other proprietary[0] and open-access[1] models (results may vary by ±2% depending on the parameters used during inference).

[0]: https://paperswithcode.com/sota/multi-task-language-understa...

[1]: https://github.com/declare-lab/flan-eval/blob/main/mmlu.py#L...

I'm excited to introduce Emu2, the latest generative multimodal model developed by the Beijing Academy of Artificial Intelligence (BAAI). Emu2 is an open-source initiative that reflects BAAI's commitment to fostering open, secure, and responsible AI research. It's designed to enhance AI's proficiency in handling tasks across various modalities with minimal examples and straightforward instructions.

Emu2 has demonstrated superior performance over other large-scale models like Flamingo-80B in few-shot multimodal understanding tasks. It serves as a versatile base model for developers, providing a flexible platform for crafting specialized multimodal applications.

Key features of Emu2 include:

- A more streamlined modeling framework than its predecessor, Emu.

- A decoder capable of reconstructing images from the encoder's semantic space.

- An expansion to 37 billion parameters, boosting both capabilities and generalization.

BAAI has also released fine-tuned versions, Emu2-Chat for visual understanding and Emu2-Gen for visual generation, which stand as some of the most powerful open-source models available today.

Here are the resources for those interested in exploring or contributing to Emu2:

- Project: https://baaivision.github.io/emu2/

- Model: https://huggingface.co/BAAI/Emu2

- Code: https://github.com/baaivision/Emu/tree/main/Emu2

- Demo: https://huggingface.co/spaces/BAAI/Emu2

- Paper: https://arxiv.org/abs/2312.13286

We're eager to see how the HN community engages with Emu2 and we welcome your feedback to help us improve. Let's collaborate to push the boundaries of multimodal AI!

- Code: https://github.com/baaivision/Emu/Emu2