**ColBERT-Zero: To Pre-train Or Not To Pre-train ColBERT models?**
•
10
AI & ML interests
None defined yet.
Recent Activity
View all activity
Papers
ColBERT-Zero: To Pre-train Or Not To Pre-train ColBERT models
LightOnOCR: A 1B End-to-End Multilingual Vision-Language Model for State-of-the-Art OCR
Articles
First large-scale fully pre-trained ColBERT model using only public data, outperforming GTE-ModernColBERT and GTE-ModernBERT
-
ColBERT-Zero: To Pre-train Or Not To Pre-train ColBERT models
Paper • 2602.16609 • Published • 5 -
lightonai/ColBERT-Zero
Sentence Similarity • 0.1B • Updated • 21 • 12 -
lightonai/ColBERT-Zero-supervised
Sentence Similarity • 0.1B • Updated • 15 • 2 -
lightonai/ColBERT-Zero-unsupervised
Sentence Similarity • 0.1B • Updated • 15 • 1
Visual long document VLMs based on Mistral-Small-3.1-24B-Instruct-2503 and Qwen3-VL-32B-Instruct
This collection gather datasets for embeddings pre-training and fine-tuning.
French language models. These model were trained in early 2021 following the then scaling laws and using the exact same training data as the CamemBERT
State-of-the-art late interaction code retrieval models
-
lightonai/LateOn-Code-edge
Sentence Similarity • 16.8M • Updated • 722 • 17 -
lightonai/LateOn-Code
Sentence Similarity • 0.1B • Updated • 94 • 14 -
lightonai/LateOn-Code-edge-pretrain
Sentence Similarity • 16.8M • Updated • 61 • 2 -
lightonai/LateOn-Code-pretrain
Sentence Similarity • 0.1B • Updated • 16 • 2
LightOnOCR-2-1B: a lightweight high-performance end-to-end OCR model family
The Case for End-to-End and Efficient Domain-Specific Vision-Language Models for OCR
Bringing BERT into modernity via both architecture changes and scaling
-
answerdotai/ModernBERT-base
Fill-Mask • 0.1B • Updated • 952k • 995 -
lightonai/GTE-ModernColBERT-v1
Sentence Similarity • Updated • 23.5k • 155 -
lightonai/Reason-ModernColBERT
Sentence Similarity • Updated • 17.3k • • 207 -
lightonai/modernbert-embed-large
Sentence Similarity • 0.4B • Updated • 43.6k • • 28
A suite of autoregressive generative models for protein sequences, with up to 1.2Bparameters, trained on over 280 million protein sequences.
900M models trained on 25BT to compare different data processing choices (filtering, sentence dedup, minhash, etc)
First large-scale fully pre-trained ColBERT model using only public data, outperforming GTE-ModernColBERT and GTE-ModernBERT
-
ColBERT-Zero: To Pre-train Or Not To Pre-train ColBERT models
Paper • 2602.16609 • Published • 5 -
lightonai/ColBERT-Zero
Sentence Similarity • 0.1B • Updated • 21 • 12 -
lightonai/ColBERT-Zero-supervised
Sentence Similarity • 0.1B • Updated • 15 • 2 -
lightonai/ColBERT-Zero-unsupervised
Sentence Similarity • 0.1B • Updated • 15 • 1
State-of-the-art late interaction code retrieval models
-
lightonai/LateOn-Code-edge
Sentence Similarity • 16.8M • Updated • 722 • 17 -
lightonai/LateOn-Code
Sentence Similarity • 0.1B • Updated • 94 • 14 -
lightonai/LateOn-Code-edge-pretrain
Sentence Similarity • 16.8M • Updated • 61 • 2 -
lightonai/LateOn-Code-pretrain
Sentence Similarity • 0.1B • Updated • 16 • 2
Visual long document VLMs based on Mistral-Small-3.1-24B-Instruct-2503 and Qwen3-VL-32B-Instruct
LightOnOCR-2-1B: a lightweight high-performance end-to-end OCR model family
The Case for End-to-End and Efficient Domain-Specific Vision-Language Models for OCR
This collection gather datasets for embeddings pre-training and fine-tuning.
Bringing BERT into modernity via both architecture changes and scaling
-
answerdotai/ModernBERT-base
Fill-Mask • 0.1B • Updated • 952k • 995 -
lightonai/GTE-ModernColBERT-v1
Sentence Similarity • Updated • 23.5k • 155 -
lightonai/Reason-ModernColBERT
Sentence Similarity • Updated • 17.3k • • 207 -
lightonai/modernbert-embed-large
Sentence Similarity • 0.4B • Updated • 43.6k • • 28
French language models. These model were trained in early 2021 following the then scaling laws and using the exact same training data as the CamemBERT
A suite of autoregressive generative models for protein sequences, with up to 1.2Bparameters, trained on over 280 million protein sequences.
900M models trained on 25BT to compare different data processing choices (filtering, sentence dedup, minhash, etc)