For our third #lovedata25 workshop @pnadelofficial.bsky.social taught about building your own search engines with NLP. He demonstrates his streamlit app where you can load your own data and customize your search. It just so happens that the LC’s crowdsourced transcriptions are perfectly formatted.

Our Love Data line-up: @kaylendwyer.bsky.social - Handwriting Text Recognition w/ @transkribus.bsky.social @akijas.bsky.social - TEI LeafWriter @pnadelofficial.bsky.social - building a search engine Finally, @douglassday.bsky.social transcribe-a-thon. We'll be using Douglass Day data all week!

I have been saying for years that 98% of the computation in LLMs is wasted, it’s just that figuring out what matters was harder than writing checks

ColPali is landed at @hf.co transformers and I have just shipped a very lean fine-tuning tutorial in smol-vision 🤠💗 QLoRA fine-tuning with 4-bit with bsz of 4 can be done with 32 GB VRAM and is very fast! ✨ github.com/merveenoyan/...

A lot has happened in DH at Tufts and the DH MA has become very exciting.

Got some open source LLMs running on the Tufts cluster, so you know I had to make a Latin finetune of Llama2-7b. Really impressive results for only 6-7 hours of training.