Ishrat Jahan (@coreprinciple_) header
Ishrat Jahan (@coreprinciple_) profile picture

LightOn

Joined April 2026
Photo by Ishrat Jahan (@coreprinciple_)Photo by Ishrat Jahan (@coreprinciple_)
Building multimodal information retrieval systems. Also: hi, I'm Ishrat. coreprinciple.vercel.app
3
225
Ishrat Jahan retweeted
We surfed the dense and late-interaction waves… now it's time to SlideOn sparse retrieval 🐕 🚀 Meet SlideOn and mSlideOn, fully open SOTA sparse retrievers at 174M and 339M parameters, trained by @LightOnIO R&D Models: huggingface.co/collections/l… Blog: huggingface.co/blog/lightona…
7
25
7
76
7,798
Ishrat Jahan retweeted
World-leading document recognition. Beyond text extraction. Introducing LightOnOCR-3, our new family of state-of-the-art AI models for document intelligence. Building on LightOnOCR-2, it recognizes text and handwriting, describes images, extracts data from charts and preserves document structure, all in a single processing step. 🏆 Leading benchmark results : ahead of Chandra-OCR-2, Mistral OCR 4.1 and dots.mocr on OlmOCR-Bench, and first among open-weight models on ParseBench. 📊 More information from every document : turn text, tables, images and charts into content your business applications can use. ⚡Compact and efficient: 0.8B and 4B versions, with document processing up to twice as fast as LightOnOCR-2. 🔓Apache 2.0 license: use, adapt and deploy the models on your own infrastructure. From handwritten administrative files to financial reports and technical documents, make more of your information searchable and usable. 👉 Explore the models and benchmark results: huggingface.co/blog/lightona… 👉🏻Explore the model page on our Website : lighton.ai/research/lightono… #LightOn #DocumentIntelligence #OCR
28
88
21
665
61,839
Drumroll please! After DenseOn and LateOn, we release mDenseOn & mLateOn: two fully open 307M retrievers for multilingual, long-context AND code search! Same backbone, same data, using translate-train. But very different behaviour once we leave the training languages 👀
🤖 Made with AI
1
13
39
1,456
"Multilingual when?" Glad to finally answer today! Introducing mDenseOn & mLateOn, the next generation of fully open models for multilingual, long-context, and code retrieval Same openness, same frontier performance, but now for everything
12
37
7
115
14,107
1/ Fun news 🎉 WanderLust won 3rd prize in the @OpenBMB category at the @huggingface × @Gradio Build Small Hackathon! A map app that asks: What if you want to explore as you go? 🗺 huggingface.co/spaces/build-…
3
1
8
957
4/ Full write-up with the orienteering solver, the zero-hallucination gate, the field notes: huggingface.co/blog/coreprin… Thanks @OpenBMB , this model punches way above its weight.
1
32
If you'd want a bigger, better WanderLust someday, tell us 👀
22
Rerankers are a key component of search pipeline With the advent of multimodal search, we released MonoQwen to fill the gap @coreprinciple_ worked hard on a long overdue update, notably by making it work on both image AND text and making it listwise!
Today, we release a collection of multimodal rerankers, handling both text & document images powered by @coreprinciple_ 's insane work! The 2B kinda mogs other same-size open rerankers 👀 (yes, there are so few of them 😭) 62.66 NDCG@10 on ViDoRe V3! Here’s what we learned 🧵
2
8
50
4,179
Ishrat Jahan retweeted
Your reranker decides what your agent gets to read. Text or image, same call. Introducing LightOn-rerank, a new collection of multimodal rerankers built to rank text passages and scanned pages on the same relevance scale. Across Pointwise and Listwise models, choose the right balance of quality, size, and serving speed for your stack. 🥇 On ViDoRe V3, LightOn-rerank-LW-4B outperforms Qwen’s 8B model at half the size. 🚀 The 2B model is faster to serve while retaining more than 97% of the quality. Less noise in context. Better evidence for your agent. 🙌 Kudos to @coreprinciple_ for this milestone! 🔗 Full breakdown: lighton.ai/lighton-blogs/int…
6
12
706
So proud of this one. My first published research post! Its a family of rerankers that handles text and document pages with a single adapter, leading every open model in its class on ViDoRe V3. Thank you @AmelieTabatta for being the best collaborator through every ablation 💛
Today, we release a collection of multimodal rerankers, handling both text & document images powered by @coreprinciple_ 's insane work! The 2B kinda mogs other same-size open rerankers 👀 (yes, there are so few of them 😭) 62.66 NDCG@10 on ViDoRe V3! Here’s what we learned 🧵
8
15
1,149
Today, we release a collection of multimodal rerankers, handling both text & document images powered by @coreprinciple_ 's insane work! The 2B kinda mogs other same-size open rerankers 👀 (yes, there are so few of them 😭) 62.66 NDCG@10 on ViDoRe V3! Here’s what we learned 🧵
🤖 Made with AI
3
14
4
49
9,605
WanderLust — our @huggingface Build Small hackathon build: A→B routes that detour through the places you'll actually love, inside a fixed time budget. Not the fastest line. App: huggingface.co/spaces/build-… Field notes: huggingface.co/blog/coreprin…
1
74
What if you could take three completely different model families… and distill them into one tiny model? 🤯 📜 Paper: arxiv.org/pdf/2605.21699 MOPD (Multi-Teacher On-Policy Distillation) has become a standard procedure in post-training. We already distill multiple specialized variants of the same model into a single set of weights. But what if we could go further - and distill models from entirely different families? Turns out, it is possible. Today we’re releasing a paper on cross-tokenizer distillation - our first steps in this exciting direction. 📄 We distilled Qwen3-4B, Phi-4-Mini, and Llama-3B into Llama-3.2-1B. MMLU jumped from 32.05 → 46.32 when using multiple teachers. 📈 The team is now working on Nemo-RL integration so the community can try this method in their own settings. Plus, we are scaling experiments up. 🚀
🤖 Made with AI
50
323
39
2,692
1,363,119
Ishrat Jahan retweeted
This is a living world you can step into, explore, and direct. 📝 OverView: happyoyster.cn/docs @AlibabaGroup @HappyHorseATH #HappyOyster #Alibaba-ATH #WorldModel
2
2
30
11,489
Ishrat Jahan retweeted
Qwen3.5 27B vs Gemma 4 31B IT  Here is why I think Gemma 4 31B wins: Benchmark scores of both models are very close. Generalization is not. I reused NVIDIA’s CoDeC metric to probe "benchmark affinity". Officially, CoDeC > 80 suggests contamination. I've never seen this. But even below that, it can still reveal how benchmark-affine a model is. Higher CoDeC scores mean the model is more "comfortable" with the benchmark. So the ideal profile is: high benchmark scores + low CoDeC. By that lens, Gemma 4 looks clearly stronger than Qwen3.5. Gemma 4 likely generalizes better to unseen tasks. I observed the same pattern earlier with Gemma 3 vs Qwen3. I wrote about this a few months ago: kaitchup.substack.com/p/did-…
23
35
15
462
96,841
🤯 GEMMA 4 + OPUS 4.6 REASONING DROPPED @kaiostephens goal: produce a Gemma 4-31B reasoning adapter trained only on Opus reasoning 🧠 What the model is: 🧬 Tiny QLoRA adapter on Gemma 4 31B-it 📊 Fine-tuned on ~1,900 curated Opus Examples ⚡ Trained in ~1 hour on a single GH200 GPU 📖 Fully open Apache 2.0 What it does: ✨ Boosts overall quality, coherence, and personality 🧮 Stronger math, code, and Opus problem solving 💬 More refined, thoughtful responses 🏠 Built for local agents, workflows, and heavy daily Vs base Gemma 4 31B: 📐 Same efficient base model, no extra size or speed 📈 Noticeable step up in real-world depth and quality 💪 Base was already strong this levels It up! Grab the adapter here 👇🏻 huggingface.co/kai-os/gemma4…
19
116
8
1,183
84,640