Daniel van Strien

OCR models often get stuck in loops, so I asked Claude Opus if it could stop one from looping by editing its weights. It changed 0.0085% of them and with no training, it now reads this 1881 page instead of writing 16,399 characters of dots! Prompt was basically: "read this, find a smallish OCR model and fix its repetition doom loops. You can spend up to 5 hours and $100 on @huggingface Jobs." It only took under 3 hours in the end and about $10 of Jobs compute.

I trained a specialist decision model for curating web-scale data. It's small (380M, fine-tuned from UltraBERT-380M), so it runs relatively cheaply at scale: 1M FineWeb pages took 19 minutes and ~$5 on @huggingface Jobs. It thinks a quarter of those pages are product pages. 6% passed a simple "educational, not spam" filter, and 1 page in 1,020,000 got "outstanding". I've shared the model, the training data (with the source, labeller, and licence for every row), and the full recipe. I hope more people share how they prepare data for training open decision models! We've had some cool open model releases. Time for some more data releases! huggingface.co/davanstrien/u…

Introducing Surya OCR 2.5! - Most accurate open OCR model under 1B params (85.9 on OmniParseBench) - 1.7x faster than Surya 2, 4.5 pages/s on H100 - 689M params, with character tokenization

Meet NVIDIA Nemotron 3.5 Transcribe 🎙️ An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint. Watch it keep track of who said what when everyone’s talking at once. 1/ 🧵

After a month or so of open Jev-style decision models, here's what I found on the question of data. Most of the findings come from one paper, "Releasing the Inner Jev" (@JonathanGeuter et al.). You may need less training than you think. Show the options as letters, run one forward pass, and read the scores for "A", "B", "C"… An instruct model done this way already gets close to Jev: Gemma 4 31B scores 79.6 on OpenDecisions vs Jev's 80.1. LoRA vs the head (31B ablations): → Train only the head, backbone frozen: 78.9 (slightly below zero-shot) → Train only LoRA, head frozen at the letter rows: 84.9 → Full recipe, LoRA + head: 84.4 So the gain comes from adapting the representations. The letter readout is already a good head. Also: → Short runs win (400–800 steps); longer runs mostly lose calibration → Smaller models gain less: no gain at 4B without extra tricks → A validation split from your own training data predicted the leaderboard badly I also checked ~100 decision-related datasets on the Hub: → Most training data is classic classification/NLI datasets recast as choice/yes-no/score questions → Much of the rest is labelled by an LLM teacher → Eval sets were tiny (100–350 questions) until OpenDecisions (2,352) → Labels need re-checking: when Contested Fast Decisions relabelled the items 8 models disagreed on, 52% changed I'm working on this a bit more to understand what data actually helps when you train one for your own task. So far I've reproduced the paper's zero-shot 4B result on @huggingface Jobs (64.7 vs 63.6, ~$0.20). Paper: huggingface.co/papers/2610.1… Datasets: huggingface.co/collections/d…

FineBooks wants to help re-OCR the public domain to unlock a rich source of data for AI, for researchers, and for enjoyment. Step one: work out which modern OCR models are actually good enough. So we built a leaderboard. 14 open models, scored on 2,165 pages of expert human transcription. The best read historical print at ~97.6% character accuracy, for under $2 per 1,000 pages. A new @huggingface and @AiEleuther project, with @storytracer! Blog post: huggingface.co/blog/finebook…