Meet SpicyLlama2: Our Improved AI Model
Added 2023-09-19 16:49:19 +0000 UTCAfter several weeks of rigorous testing, we are proud to announce that our first custom model, trained using our proprietary data, is now becoming our primary model. This model has been shaped and improved based on the labeled data we gather each time you provide feedback on the responses.
Introducing SpicyLlama2, a refined version of Llama2-13B. We've been running this model on a random selection of generation requests to accumulate rating data. With data from over 1.2 million labeled responses, SpicyLlama2 has surpassed our default model, registering 5% fewer negative feedbacks. While this improvement might seem modest, it signifies a promising step forward.
Looking ahead, one of our key objectives is to train our models on an even larger dataset of labeled feedback.
Comments
Great, thanks!
Supahalex
2024-03-27 17:07:33 +0000 UTCThis is a nice outlet, to say things I never would in real life. Thank you, creators.
David G
2023-10-01 20:07:47 +0000 UTCOur decision to switch was based on quantitative data, next week we'll run a user survey asking for feedback on the new model and the results will help determine the plan moving forward. Because of the volume of generation requests and the scarce availability of powerful GPUs, it's very hard for us to consider offering multiple models to choose from, all our resources are going toward supporting the growth in daily users.
Karl Bernard
2023-09-22 13:35:35 +0000 UTC
