Documentation of the Kick-off session on 02/10/2024
(Sub) Working Group 2 /Task Team on LLM Implementation and LLM Services (E-AI)
Participants:
Agenda:
Onboarding to the Working Group 2 and notice to Terms of Reference document
Agree on a shared long term vision and targets for the WG inspired by KNMI E-AI meeting
Agree on short term (~2-3 months) feasible deliverables (we will propose some ideas, you will be prompted to complement with your own)
Agree on how we organize work to achieve these objectives (weekly agile like KanBan/sprint planning + github repo ?)
Ongoing Projects using LLMs participants would like to share (collaborative table)
Long term goals and use-cases with LLMs (collaborative table)
Expectations from this WG on LLMs (collaborative table)
Suggestions for this WG
Dedicated session on cybersecurity and privacy including service provider and IT department.
FG, KS, HF, NR, BP
Workshop on techniques and technical solutions for adapting LLMs to your data without total retraining: prompt engineering/injection, RAG, Lora, Qlora, PEFT, …
JS, FG,NR, JC, KK, FH, BP, HF
* RAG: Retrieval-Augmented Generation (not necessarily mean fine-tuning)
* PEFT: Parameter Efficient Fine Tuning
* LoRA: Low-Rank Adaptation (a type of PEFT)
* QLoRA: LoRA for quantized LLMs
Collect field and experimental feedback on real use-cases related to our domain on the various techniques mentioned above.
Workshop on small language/Multimodal Models : smaller architectures specialized in weather applications
Dedicated Workshop on Marine Weather reports - sharing code ? FG, KS, NR, JC, FH, BP
Focus on RAG of Foundation Models/ Large/Small Language Models, which are already pre-trained instead of own training or fine-tuning (resource-intensive)
Advice on technical setups? Both cloud solutions and on-prem access to GPUs.
Compliance, Validation, Testing and Ethical/Legal Questions (WG8 Ethics?) … KS, NR
Short term Deliverables & Next Steps
Participants and dates for 3 workshops in 2024 (date finding with doodle-like tool with participants with interest in participation, all participatants of first session are invited to join the 3 workshops):
Cybersecurity and privacy including service provider and IT department. [FG, KS, HF, NR, BP]
Techniques and technical solutions for adapting LLMs to your data without total retraining (prompt engineering/injection, RAG, Lora, Qlora, PEFT) [JS, FG,NR, JC, KK, FH, BP, HF]
Dedicated Workshop on Marine Weather reports – sharing code [FG, KS, NR, JC, FH, BP]
Create a shared Rag github project incl. a cookbook on a real use case? Incl links to good tutorials in the README.md. Repo should include all dependencies to have a reproducible setup. FG, HF, JS, NR, KK, JC, KrKr
Create a shared github project for text products generation: marine, TAFs, … FG, JS, NR, KS, KK, JC, KrKr
Invite speakers from tech companies (re the conversation on use of ChatGPT) FG,