Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
Current trends + advances in Open LLMs - brief roundup by DiscoResearch
Overview of recent open‑LLM advances, demo of DiscoLM German, lessons from large‑scale fine‑tuning projects, and practical guidance for production deployment, covering data‑filtering and DPO techniques.
Capability advances over the last few weeks/months enable LLM applications with local / on-premise models, that werent possible until a few weeks/months ago.
We want to highlight a few of them with a short demo of DiscoLM German (successor to EM German, the most-downloaded open german-focussed LLM; will be released end of 2023).
We´d also like to talk about lessons learned from recent DiscoLM projects (DiscoLM 120b, #3 HF leaderboard + DiscoLM Mixtral, first working Mixtral finetune wordwide) and about trends, datasets + roadmap for open LLM finetuning+customization (LLM-based eval for data filtering, DPO, finetuning for Agent/RAG capabilities …) and how to get open LLMs into production.
(this is a first braindump, exact focus of the talk tbc)
DiscoResearch develops German Llama3 LLMs via continued pretraining and quantization.
EM German finetunes Llama2/Mistral/LeoLM for German text generation/retrieval.
Compose Email
Loading recent emails...