Eagle-7B: Soaring past Mistral-7B Across 100+ Languages (AI News)
Introducing Eagle 7B: The Future of Language Processing! Discover how Eagle 7B, utilizing advanced Recurrent Neural Networks (RNNs), outperforms the renowned Mistral 7B in linguistic capabilities across over 100 languages. Witness the revolution in AI efficiency as Eagle 7B achieves faster training times and quicker inference, setting a new benchmark in the field. Join us in exploring the groundbreaking potential of Eagle 7B and its impact on the world of language processing!
RWKV Announcement: https://blog.rwkv.com/p/eagle-7b-soaring-past-transformers
HuggingFace Demo: https://huggingface.co/spaces/BlinkDL/RWKV-Gradio-2
BlinkDL Twitter (follow them!): https://twitter.com/BlinkDL_AI
Intro video credit: Nper eth
by Ai Flux
linux foundation
Canadian here.. working in a french environment and thinking in English
Run the write a poem praising Joe Biden/Donald Trump test. I did.
If this checks out, it'll revolutionary. Everybody buried RNNs, but there might be life in them yet. At very least for specialized purposes, or along transformers.
I've known a number of people who have been using RWKV for more than a year now on problems like the ARC challenge – they are all very enthusiastic about it.
Nice to see Japanese on the list – Google translate is garbage with Japanese.
If this thing scales that way, then it should be a perfect candidate as a coding assistant LLM. Looking forward to eagle-coder:13b that I can run beside my IDE on my 16 GB M1 MBP (in Ollama).
Uh-oh, you know its going to be garbo when they have to point out how “green” it is.
I don't understand how this model can be rated so high. I tried it out on HF spaces and it doesn't seem to be able to do even the simplest of tasks. I tested using it for RAG and it just prints random results, even with temperature and top p at 0. For instance, it can't tell the following except has nothing to do with the query:
—
User: Read the following Product Manual Excerpt. Tell if this excerpt says that this product has the feature "steering reset".
Product Manual Excerpt: ABS Bleeding
The ABS Bleeding function is available for Chrysler, GM, Hyundai/Kia,
Mazda and Toyota vehicles only. Procedures vary between vehicle
makes and models.
If an error occurs while performing ABS bleeding procedures,
an “advisory” message displays. Choose Exit or Back , as
necessary, to return to the Service Reset menu.
Assistant:
—-
Yes, the product manual excerpt states that this product has the feature "steering reset".
More experts like you needed here! Thanks for the overview!
Guys, keep in mind this is a foundational model and it isn't finetuned for instruction following. So it requires heavy prompting or finetuning to output meaningful stuff.
I don't know if it's worth it. Eg. I was using wizard supercot storry teller 30b. I changed this to super "carbonvilan 10.4b" maybe it have a lot of more knowledge and vocabulary, but very often halucinates and it loses the thread. For this model it would be simmilar. Gpt playground is better and almost uncensored
It hallucinates continuously, doesn't answer what I asked. Jumps between topics even within a reply. Unusable.
god bless,thanks for the information
I like how you use real-world examples to illustrate your points, it makes it easier to relate.
it would be nice to see how much ram the model uses too as a comparison especially as we march to portable systems like rpi5
Im regulary test llms in Russian, i will try this one for sure
No way this is coming anywhere close to Mistral quality; tested it, it's complete rubbish. Bait video.
Great overview man! Keep up the good work
4:12 "Infinitive context length"?! 🤯
I have tried the demo few times but most of the output I get make no sense (and can be pretty rude)
Can this be run via Ollama or what’s the best way to dive in?
No way! I generally use Mistral 7B