Indico avec slides: https://indico.in2p3.fr/event/31207/

Notes par Michael Dell’aiera

Notes

On the Huggingface leaderboard are displayed the most relevant LLM based on the users feedback
Open source models (like Mixtral) are climbing the ladder
Sometimes, newer versions perform worse than former versions

Some kind of usage:

• Open source model for concept description in English (same usage as ChatGPT)
• LLM for astronomy: learn from the ELOG to create a debugging tool, but some questions arise, such as should we retrain the model every time a new error occurs ?
• Provide image and generate latex code
• LLM on MUST in the case of private data
• Some online application:
• Scopus IA launched by Elsevier for automatic reviewing of papers
• consensus
• perplexity (also a chrome extension) to ask question / summarize the web page
• scispace: litterature review and summarizing, etc…
• LLM to detect fake/propaganda data ?
• LLM that takes notes, transcript what we say (whisper: audio to text), but challenging because on Teams the functionnality is not working very well and it is hard to detect who is speaking, but privacy-wise it would be good to have one local for the lab
• Table Extractor : Extract info from images (like tables)

6G flagship, call from university of Oulu, interested in LLM

• how people see the future of LLMs, like applicaton areas, we can write an abstract of what we say today (deadline 31 of Jan) to be publshed in April

Custom GPT (in the Explore GPTs section), could be used for gammalearn

Call for contributions

Call for contributions to work on LLMs

See https://indico.in2p3.fr/event/31207/attachments/82398/121872/Call%20for%20Contributions.pdf

Contact Mickael Bettinelli mickael.bettinelli@univ-smb.fr if interested

Next meeting

Feb. 20, 2024