Pinned
Convert any PDF into a fillable form with 🪄AI magic. I built a free tool that: (a) takes seconds, (b) outperforms all the commercial solutions, and (c) is open source!
00:00
Joe Barrow
1,095 posts
Joe Barrow
@barrowjoseph
- If you've got 308GB to spare and want a high-quality document dataset to add to your VLM tasks, CommonForms is now hosted on @huggingface! It only took 24 hours to upload from my home network. 😭 huggingface.co/datasets/jbarr…
- This is such a good paper. I love NLP error analysis papers -- @chrmanning and co do them so well (another great example will always be "Part-of-Speech Tagging from 97% to 100%").
- Replying to @barrowjoseph
- With quantization support, @awnihannun and MLX are on fire! Here's Phi-2 running on my 8GB (!!) M2 in real time. And Mistral-7B should also work.
00:00 - Alright @Dorialexander, the Qwen2.5 VL localization is actually insanely good. This is 0-shot detection/localization/classification of checkboxes in a random document image I found online (a typically difficult task given the small resolution of the objects).
- ✅ Dissertation ✅ Defense ◻️ Sleeping for the next 24 hours All done at @ClipUmd 🎉! Thanks for everything!
- MLX is great! The Python code base is concise and high quality, making it easy to understand and extend. And it’s *fast* — I didn’t add timing information yet, but on my M2 MacBook Air, BERT is notably faster than the torch equivalent (running on its CPU). A very Apple product.Just merged a BERT example in MLX thanks to @barrowjoseph github.com/ml-explore/mlx… Easy to use + fast + flexible 🚀
- Replying to @barrowjoseph @allen_ai and @nlpnoahIf you’re unfamiliar with AllenNLP but are a beginning NLP researcher or interested party, I have a “Learn AllenNLP The Hard Way” tutorial on my site: jbarrow.ai/allennlp-the-h…
- Hey @awnihannun, I've now got an MLX implementation of phi-2 working! Missing the generate function rn, but will put together a PR tonight or tomorrow morning. Free to help make it faster?
- Replying to @barrowjosephThe neat part is that this is built on @modal, so it's fast, robust, and ~free for me to run. (👀 @charles_irl)
- Updated "AllenNLP The Hard Way" to use the newly released @ai2_allennlp 1.0! Link: jbarrow.ai/allennlp-the-h… Github: github.com/jbarrow/allenn…
- Mission accomplished and model merged, thanks to @awnihannun and @jbochi! And it runs on my M2 Air with 8GB of RAM! 🎉
00:12Phi-2 in MLX github.com/ml-explore/mlx… Super high quality 2.7B model from Microsoft that runs efficiently on devices with less RAM - Replying to @simonwFor readability, I've been throwing a `| fold -s -w 100` to the end of this to get nice, readable line breaks (e.g., screenshot is the function I have for Mistral)
![curl -s --location "https://api.mistral.ai/v1/chat/completions" \
--header 'Content-Type: application/json' \
--header 'Accept: application/json' \
--header "Authorization: Bearer $MISTRAL_API_KEY" \
--data '{
"model": "mistral-medium",
"messages": [{"role": "user", "content": "'"${prompt}"'"}]
}' | jq -r '.choices[0].message.content' | fold -s -w $width](/web/20150905204808/https://developer.android.com/samples/BasicMediaDecoder/res/values/base-https-pbs.twimg.com/media/GBQHgLQWUAAFn0C.jpg)
