
Research scientist @Numind_ai - NLG - PhD from @Sorbonne_Univ_
Joined November 2021
- Tweets98
- Following461
- Followers182
- Likes419
PP-OCRv6 processing this image in 3.3secs on a H100, gotta say I'm truly impressed. v5 was already great, I now see almost no error, shoutout @PaddlePaddle team
39
So happy to finally release NuExtract3! 🙌
A 4B open-weight model with an unmatched performance-to-size ratio for structured extraction and OCR/markdown.
Stay tuned for the Pro version, and future open-source resources for structured extraction evaluation!
We are releasing NuExtract3, a 4B open-source (Apache 2.0) reasoning OCR & Structured Extraction VLM 🧠📷📄
NuExtract3 is the first reasoning VLM specialized in both OCR (converting PDFs/Scans/Spreadsheets into Markdown) and structured extraction (document to JSON via a schema).
We trained NuExtract3 from Qwen3.5 4B via an SFT and an RL phase to provide its reasoning abilities which can be turned on and off. We find that NuExtract3 outperforms similar size models (<30B) in both the structured and OCR tasks, including specialized OCR models, making it the new reference for open-source document extraction😀.
Congrats Alexandre Constantin, Sören Dréano, and @NathanFradet for this model! Looking forward to the Pro version :)
Available on @huggingface. Links in the reply.
1
64
Nathan Fradet retweeted
We are releasing NuMarkdown-8B-Thinking, an open-source (MIT License) reasoning OCR model 🧠✨📄
NuMarkdown-8B-Thinking is apparently the first (!) reasoning VLM specialized in converting PDFs/Scans/Spreadsheets into Markdown files (typically used for RAG applications).
It generates thought tokens to figure out the layout of the document before generating the Markdown file. It is particularly good at understanding documents with weird layouts and complex tables. The number of thinking tokens can vary from 20% to 500% of the final answer, depending on task difficulty.
This model is a fine-tune of @Alibaba_Qwen 2.5-VL-7B on synthetic Doc → Reasoning → Markdown examples, followed by an RL phase (GRPO) with a layout-centric reward.
We find that NuMarkdown is outperforming generalist non-reasoning models like GPT-4o, specialized OCR models like OCRFlux, and is competitive against large reasoning closed-source models like Gemini 2.5!
Congrats to Alexandre Constantin for this tour de force
Nathan Fradet retweeted
📢NuExtract 2.0 & the NuExtract platform are OUT!📢
NuExtract 2.0 is a VLM specialized in extracting structured information (JSON output) from texts/PDFs/scans, and it is outclassing frontier LLMs at it
Nathan Fradet retweeted
Very happy and proud to present 3 late breaking demo papers this year at @ISMIRConf in November 😍 we'll present two fun tools that lower the barrier of entry for symbolic music research and an awesome new audio dataset 🧵1/n
Nathan Fradet retweeted
🎵🪇 Expressive Generation and Variations on Symbolic Music 🎶🎻
So excited to announce that the first 🍋@lemonaidexmusic research paper, which received a 'strong accept' at ISMIR is now out on Arxiv! Link below, and more details in this short 🧵
arxiv.org/abs/2410.02060
youtu.be/syQMTZyzqcg
This video is spot-on one of the biggest problem of architecture this last century. Future generations will look down on us if we continue to build without considering people’s preferences.
Thank you @aesthetic_city for your efforts in raising awareness.
Nathan Fradet retweeted
Introducing Symbolic Jazz Standards! It's a dataset of well-known jazz tunes I transcribed from the audio to the symbolic domain stem by stem (think .wav -> .midi). It's one of the largest datasets of this kind that I know of.
Symusic now only handles MIDI files, but it aims to support all symbolic music file formats, especially abc and musicXML, and provide many useful features to process data!
1
111
@RealYikaiLiao and lzqlzzq have been very effective and reactive in developing symusic and minimidi. Again, thank you for all your work which is a very appreciated contribution to the open-source and scientific community, which I have no doubt will be useful to many!
1
99
Nathan Fradet retweeted
I'm looking for participants for a listening study on #AI generated tablatures in the style of progressive metal 🤘please DM if you'd like to know a bit more about it and maybe take part 🙏
Nathan Fradet retweeted
Does any one know lawyer in Italy and could reference me to? For visa refusals, the embassy told me to find lawyers in Italy. I would like to pay whatever I can to get it sorted…
I'm delighted to announce that our paper "Byte Pair Encoding for Symbolic Music" has been accepted at #EMNLP2023 ! 🎉
Paper: arxiv.org/abs/2301.11975
Code + 🤗models: github.com/Natooz/BPE-Symbol…
Companion website: natooz.github.io/BPE-Symboli…
BPE is directly implemented in Miditok, allowing you to easily benefit from all its advantages. It is backed by 🤗tokenizers and is super fast, the tokenization / detokenization times are only slightly impacted!
github.com/Natooz/MidiTok
1
1
160
I'll see you at Singapore for #EMNLP, come meet me to talk about music, NLP or just have a drink! 🍹
1
232
















