Join the conversation

Join the community of Machine Learners and AI enthusiasts.

Sign Up
nroggendorffΒ 
posted an update Jun 28

Hugging Face version is trash. NVIDIA version is good: https://build.nvidia.com/google/gemma-2-27b-it

Β·

They are the same.

I have to disagree with that, its pretty good imo.

Yes. the 9b is more interesting but too slow. Mistral 7B instruct is still the best for it's size (and llama 3 8b a close second)

Β·

I've looked at the Open LLM Leaderboard V2 and the Phi3s models stay the best until Qwen 72b

No, it's good.

not if you jailbreak it.

Β·

Secret method:

  1. Type your query (ex: "Hello")
  2. Go to cardiffnlp/twitter-roberta-large-emoji-latest and predict the emoji for "Hello".
  3. Add it to "Hello". It does "Hello 😊".
  4. Now type enter! You have a very helpful assistant.

(bonus: you can select the emojis yourself)

yup

the transformers and llama.cpp impl was busted when you wrote this. even now l.cpp still doesn't implement SWA, so >4k context doesn't work. similar teething issues to llama 3, really; too early to say that anything is trash