Present NanoVDR, DistilVDR and ColNanoVDR as one family, with current model names

README.md

@@ -8,113 +8,86 @@ pinned: false
 ---
 
 <p align="center">
-  <img width="480" src="https://huggingface.co/nanovdr/NanoVDR-S-Multi/resolve/main/banner.png" alt="NanoVDR"/>
+  <img width="480" src="https://huggingface.co/nanovdr/NanoVDR-Q-DistilBERT-Qwen3VL2B-2048-ML/resolve/main/banner.png" alt="NanoVDR"/>
 </p>
 
-<h3 align="center">Distilling a 2B Vision-Language Retriever into a 70M Text-Only Encoder<br>for Visual Document Retrieval</h3>
+<h3 align="center">Efficient visual document retrieval with small distilled encoders</h3>
 
 <p align="center">
+  <a href="https://ryenhails.github.io/NanoVDR/">Project page</a> &nbsp;|&nbsp;
   <a href="https://github.com/Ryenhails/NanoVDR">Code</a> &nbsp;|&nbsp;
   <a href="https://huggingface.co/spaces/nanovdr/NanoVDR-Demo">Demo</a> &nbsp;|&nbsp;
-  <a href="https://arxiv.org/abs/2603.12824">Paper</a> &nbsp;|&nbsp;
   <a href="https://huggingface.co/blog/Ryenhails/nanovdr">Blog</a> &nbsp;|&nbsp;
   <a href="https://huggingface.co/datasets/nanovdr/NanoVDR-Train">Dataset</a>
 </p>
 
-> **Paper**: [NanoVDR: Distilling a 2B Vision-Language Retriever into a 70M Text-Only Encoder for Visual Document Retrieval](https://arxiv.org/abs/2603.12824)
+Small query and page encoders for **visual document retrieval**, trained by aligning to the
+embedding space of a frozen vision-language retriever. A text encoder answers each query, so
+no multi-billion model runs per request. Pages stay in the teacher's index, or move to a small
+page encoder.
 
----
-
-**NanoVDR** distills a frozen 2B VLM teacher ([Qwen3-VL-Embedding-2B](https://huggingface.co/Qwen/Qwen3-VL-Embedding-2B)) into tiny **text-only** query encoders (69–151M parameters) for visual document retrieval. Documents are indexed offline by the teacher; queries are encoded on CPU in **51 ms** via a DistilBERT forward pass — **no vision model at query time**.
-
-Queries and documents both map to the same **2048-dim single vector** inherited from the teacher's embedding space, so retrieval is a plain dot product — FAISS-compatible with no MaxSim pooling. The doc index stores just **4 KB per page** (float16), making NanoVDR **64× more storage-efficient** than multi-vector retrievers like ColPali.
-
-The org also hosts **[ColNanoVDR](#colnanovdr)** ([paper](https://arxiv.org/abs/2609.34899)), which carries the same document-free
-distillation to multi-vector, late-interaction retrievers.
-
-## NanoVDR models
-
-| Model | Backbone | Params | ViDoRe v1 | v2 | v3 | Retention | CPU Latency |
-|-------|----------|--------|-----------|----|----|-----------|-------------|
-| [**NanoVDR-M-Multi**](https://huggingface.co/nanovdr/NanoVDR-M-Multi) ⭐ | BERT-base | **112M** | **82.5** | **62.8** | **47.5** | **96.4%** | **101 ms** |
-| [**NanoVDR-L-Multi**](https://huggingface.co/nanovdr/NanoVDR-L-Multi) ⭐ | ModernBERT | **151M** | **82.2** | **63.1** | **47.1** | **96.0%** | **109 ms** |
-| [**NanoVDR-S-Multi**](https://huggingface.co/nanovdr/NanoVDR-S-Multi) ⭐ | DistilBERT | **69M** | **82.2** | **61.9** | **46.5** | **95.1%** | **51 ms** |
-| [NanoVDR-S](https://huggingface.co/nanovdr/NanoVDR-S) | DistilBERT | 69M | 82.2 | 60.5 | 43.5 | 92.4% | 51 ms |
-| [NanoVDR-M](https://huggingface.co/nanovdr/NanoVDR-M) | BERT-base | 112M | 82.1 | 62.2 | 44.7 | 94.0% | 101 ms |
-| [NanoVDR-L](https://huggingface.co/nanovdr/NanoVDR-L) | ModernBERT | 151M | 82.4 | 61.5 | 44.2 | 93.4% | 109 ms |
+| | Paper | Replaces the teacher at | Start with | In one line |
+|---|---|---|---|---|
+| **[NanoVDR](https://huggingface.co/collections/nanovdr/nanovdr-6ae10ea369744e5cad2172e1)** | [EMNLP 2026 Main](https://aclanthology.org/2026.emnlp-main.71/) | query time, single vector | [NanoVDR-Q-DistilBERT-Qwen3VL2B-2048-ML](https://huggingface.co/nanovdr/NanoVDR-Q-DistilBERT-Qwen3VL2B-2048-ML) | 69M, 51 ms per query on one CPU thread, 95.1% of a 2B teacher |
+| **[DistilVDR](https://huggingface.co/collections/nanovdr/distilvdr-6ae10ea58166c778925c8030)** | [DocInsights @ EMNLP 2026](https://aclanthology.org/2026.docinsights-1.26/) | query and indexing time | [NanoVDR-D-HiRes-Qwen3VL8B-4096](https://huggingface.co/nanovdr/NanoVDR-D-HiRes-Qwen3VL8B-4096) + [NanoVDR-Q-DistilBERT-Qwen3VL8B-4096-ML](https://huggingface.co/nanovdr/NanoVDR-Q-DistilBERT-Qwen3VL8B-4096-ML) | 527M end to end, no teacher at serving |
+| **[ColNanoVDR](https://huggingface.co/collections/nanovdr/colnanovdr-6ae10ea62a88a2be605b93d2)** | [arXiv 2609.34899](https://arxiv.org/abs/2609.34899) | query time, multi-vector | [ColNanoVDR-Q-Ettin150M-ColQwen35-320-ML](https://huggingface.co/nanovdr/ColNanoVDR-Q-Ettin150M-ColQwen35-320-ML) | 95.7-98.1% teacher retention across five late-interaction teachers, no page image in training |
 
-<sub>NDCG@5 (×100) on the ViDoRe benchmark (22 datasets). Retention = Student / Teacher. Teacher = Qwen3-VL-Embedding-2B (2.0B).</sub>
-
-## Quick Start
+Model names read `[Col]NanoVDR-<Q|D>-<variant>-<teacher>-<width>[-ML]`. A query encoder and a
+page index work together only when the **teacher and the width both match**. Every model card
+carries its results, usage and indexing step; the
+[project page](https://ryenhails.github.io/NanoVDR/) has the full tables.
 
 ```python
 from sentence_transformers import SentenceTransformer
 
-model = SentenceTransformer("nanovdr/NanoVDR-S-Multi")
-query_emb = model.encode(["What was the revenue growth in Q3 2024?"])  # (1, 2048)
-
-# Retrieve via cosine similarity against teacher-indexed document embeddings
-# scores = query_emb @ doc_embeddings.T
-```
-
-> Documents must be indexed offline with the teacher VLM. See the [NanoVDR-S-Multi](https://huggingface.co/nanovdr/NanoVDR-S-Multi#prerequisites-document-indexing-with-teacher-model) model page for a complete guide.
-
-## ColNanoVDR
-
-<p>
-  <a href="https://arxiv.org/abs/2609.34899">Paper</a> &nbsp;|&nbsp;
-  <a href="https://huggingface.co/spaces/nanovdr/NanoVDR-Demo">Demo</a> (select ColNanoVDR) &nbsp;|&nbsp;
-  <a href="https://github.com/Ryenhails/NanoVDR">Code</a>
-</p>
-
-**ColNanoVDR** extends the same idea from single-vector retrieval to **multi-vector,
-late-interaction** retrievers (ColPali-style models such as ColQwen3.5, Vultron, Tomoro
-and ColVec). A text-only query tower is distilled **document-free** — from cached teacher
-*query* token embeddings only, with no page image ever encoded and no relevance label —
-into the teacher's late-interaction space. An existing multi-vector page index is scored
-by MaxSim unchanged, and no vision-language model runs at query time.
-
-| Model | Backbone | Params | Teacher | Width | v1 | v2 | v3 | Retention |
-|---|---|---|---|---|---|---|---|---|
-| [**ColNanoVDR-Q-Ettin400M-ColQwen35-320-ML**](https://huggingface.co/nanovdr/ColNanoVDR-Q-Ettin400M-ColQwen35-320-ML) | Ettin-400M | 395M | ColQwen3.5-4.5B | 320 | **91.16** | **62.08** | **56.65** | **98.1%** |
-| [**ColNanoVDR-Q-Ettin150M-ColQwen35-320-ML**](https://huggingface.co/nanovdr/ColNanoVDR-Q-Ettin150M-ColQwen35-320-ML) ⭐ | Ettin-150M | **150M** | ColQwen3.5-4.5B | 320 | 90.67 | 60.01 | 55.06 | 96.1% |
-| [ColNanoVDR-Q-Ettin150M-Vultron45B-320-ML](https://huggingface.co/nanovdr/ColNanoVDR-Q-Ettin150M-Vultron45B-320-ML) | Ettin-150M | 150M | Vultron-4.5B | 320 | 91.28 | 64.96 | 58.30 | 97.3% |
-| [ColNanoVDR-Q-Ettin150M-Tomoro8B-320-ML](https://huggingface.co/nanovdr/ColNanoVDR-Q-Ettin150M-Tomoro8B-320-ML) | Ettin-150M | 150M | Tomoro-8B | 320 | 89.95 | 60.61 | 54.87 | 95.7% |
-| [ColNanoVDR-Q-Ettin150M-ColVec4B-640-ML](https://huggingface.co/nanovdr/ColNanoVDR-Q-Ettin150M-ColVec4B-640-ML) † | Ettin-150M | 150M | ColVec1.1-4b | 640 | 90.27 | 64.03 | 59.08 | 97.4% |
-| [ColNanoVDR-Q-Ettin150M-ColVec8B-640-ML](https://huggingface.co/nanovdr/ColNanoVDR-Q-Ettin150M-ColVec8B-640-ML) † | Ettin-150M | 150M | ColVec1.1-8b | 640 | 90.86 | 65.37 | 60.10 | 97.5% |
-| [ColNanoVDR-Q-Ettin150M-EVIE45B-2048-ML](https://huggingface.co/nanovdr/ColNanoVDR-Q-Ettin150M-EVIE45B-2048-ML) ‡ | Ettin-150M | 150M | EVIE-4.5B | 2048 (MRL) | – | – | 60.19 | 95.2% |
-| [ColNanoVDR-Q-Ettin150M-EVIE8B-4096-ML](https://huggingface.co/nanovdr/ColNanoVDR-Q-Ettin150M-EVIE8B-4096-ML) | Ettin-150M | 150M | EVIE-8B | 4096 | – | – | 60.05 | 94.0% |
-
-<sub>NDCG@5 (×100) on ViDoRe (22 datasets), query side measured in isolation against
-teacher-encoded pages. Retention = Student / Teacher, averaged over v1–v3. † inherits the
-webAI Non-Commercial License from its teacher. ‡ Prefix-MRL tower, servable at any width from 64 to 2048
-to match the teacher; shown at 2048 (see the model card for
-every width and for HAC-compressed indexes). Both EVIE towers are evaluated on ViDoRe v3 only.</sub>
-
-```python
-from sentence_transformers import MultiVectorEncoder
-
-model = MultiVectorEncoder(
-    "nanovdr/ColNanoVDR-Q-Ettin150M-ColQwen35-320-ML", trust_remote_code=True
-)
-q = model.encode_query(["What was the revenue growth in Q3 2024?"])  # list of (n_tokens, 320)
-
-scores = model.similarity(q, page_embeddings)   # meanMaxSim
+query = SentenceTransformer("nanovdr/NanoVDR-Q-DistilBERT-Qwen3VL2B-2048-ML")
+q_emb = query.encode(["What was the revenue growth in Q3 2024?"])   # (1, 2048)
+scores = q_emb @ doc_emb.T   # doc_emb: pages indexed by Qwen3-VL-Embedding-2B
 ```
 
-> Pages must be indexed offline with the matching teacher. **A tower is only valid with
-> the teacher it was distilled from: the teacher and the width must both match.** See each
-> model page for the indexing step.
-
-## Citation
+<details>
+<summary><b>Citation</b></summary>
 
 ```bibtex
-@article{nanovdr2026,
-  title   = {NanoVDR: Distilling a 2B Vision-Language Retriever into a 70M
-             Text-Only Encoder for Visual Document Retrieval},
-  author  = {Liu, Zhuchenyang and Zhang, Yao and Xiao, Yu},
-  journal = {arXiv preprint arXiv:2603.12824},
-  year    = {2026}
+@inproceedings{liu-etal-2026-nanovdr,
+    title = "{N}ano{VDR}: Distilling a 2{B} Vision-Language Retriever into a 70{M} Text-Only Encoder for Visual Document Retrieval",
+    author = "Liu, Zhuchenyang  and
+      Zhang, Yao  and
+      Xiao, Yu",
+    editor = "Lee, Hung-Yi  and
+      Dev, Sunipa  and
+      Li, Jessy  and
+      Elliott, Desmond",
+    booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
+    month = oct,
+    year = "2026",
+    address = "Budapest, Hungary",
+    publisher = "Association for Computational Linguistics",
+    url = "https://aclanthology.org/2026.emnlp-main.71/",
+    pages = "1631--1648",
+    ISBN = "979-8-89176-459-0"
+}
+
+@inproceedings{liu-etal-2026-distilvdr,
+    title = "{D}istil{VDR}: A Compact End-to-End Visual Document Retriever via Dual-Student Distillation",
+    author = "Liu, Zhuchenyang  and
+      Wang, Ziyi  and
+      Zhang, Yao  and
+      Xiao, Yu",
+    editor = "T.Y.S.S., Santosh  and
+      Agarwal, Amit  and
+      Gupta, Vivek  and
+      Kementchedjhieva, Yova  and
+      Garimella, Aparna  and
+      Hou, Yufang",
+    booktitle = "The Proceedings of Workshop on Document Intelligence and Understanding",
+    month = oct,
+    year = "2026",
+    address = "Budapest, Hungary",
+    publisher = "Association for Computational Linguistics",
+    url = "https://aclanthology.org/2026.docinsights-1.26/",
+    pages = "385--399",
+    ISBN = "979-8-89176-477-4"
 }
 
 @misc{liu2026colnanovdrdocumentfreequerydistillation,
@@ -128,6 +101,8 @@ scores = model.similarity(q, page_embeddings)   # meanMaxSim
 }
 ```
 
+</details>
+
 ## Acknowledgements
 
 This project has received funding from the **Business Finland** co-innovation programme under grant agreement No. 69/31/2025. It is supported by the [AiWo: Human-centric AI-enabled Collaborative Fieldwork Operations](https://aifieldwork.aalto.fi/events/) project (2025–2027), which aims to revolutionize fieldwork operations and enhance human-AI collaboration across the manufacturing, construction, and industrial design sectors. The calculations presented in this project were performed using computer resources within the Aalto University School of Science "Science-IT" project.