nanovdr/NanoVDR-Q-DistilBERT-Qwen3VL8B-4096-ML

NanoVDR-Q-DistilBERT-Qwen3VL8B-4096-ML — the query tower for the 4096-d space

Project page  |  Models  |  Code  |  Paper

---

A 70M text-only encoder that maps a query into the embedding space of Qwen3-VL-Embedding-8B, distilled by pointwise cosine alignment against that teacher's cached query embeddings. The -ML suffix marks the multilingual training mixture: 711K English queries plus 778K MarianMT translations into five Latin-script European languages.

Any half of the teacher, swapped

Both towers land in the same frozen teacher's embedding space, so either is a drop-in replacement for the corresponding half of that teacher. All four combinations retrieve; they differ in what you no longer have to run:

_{Average NDCG@5 over ViDoRe v1+v2+v3, this model paired with NanoVDR-D-HiRes-Qwen3VL8B-4096, against Qwen3-VL-Embedding-8B.}

The document towers in this space:

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("nanovdr/NanoVDR-Q-DistilBERT-Qwen3VL8B-4096-ML")
q = model.encode(["What was the revenue growth in Q3 2024?"])   # (1, 4096), L2-normalised

Retrieval is a dot product against page vectors from a matching document tower:

from transformers import AutoModel, AutoImageProcessor

doc = AutoModel.from_pretrained("nanovdr/NanoVDR-D-HiRes-Qwen3VL8B-4096", trust_remote_code=True).eval()
proc = AutoImageProcessor.from_pretrained("nanovdr/NanoVDR-D-HiRes-Qwen3VL8B-4096", trust_remote_code=True)
d = doc.encode(pages, proc, batch_size=4)

scores = q @ d.T

Do not add an instruction prefix

Pass the raw query. The teacher's targets were cached with an instruction, but the student was trained to reproduce those targets from the bare query text, and every number below was measured that way. Prepending an instruction at inference moves the input off the distribution the tower was fitted on.

Results

NDCG@5 on ViDoRe. Query side in isolation, scored against teacher-encoded pages, which isolates this tower from document-side error:

End to end with the document towers, no teacher anywhere:

Query encoding costs 2.07 ms per query at batch 1 on an H200, including tokenisation. There is no vision model in the query path.

Backbone ablation

Measured under query-side isolation, all distilled from the same teacher:

ModernBERT is 1.04 points better for 2.2x the parameters and 5.1x the latency, which is why the small one ships as the default.

Training

Peak LR is half what the document tower uses; at 1e-3 a freshly initialised 768→4096 head diverges at the peak of the one-cycle schedule.

Reproducibility

Re-encoding the 500 ViDoRe arxivqa queries with this package and scoring against the cached teacher page embeddings:

NDCG@5  teacher queries x teacher pages : 86.91
NDCG@5  student queries x teacher pages : 83.77   (96.4% retention)

which matches our internal evaluation for that dataset exactly.

Limitations

• English and five Latin-script European languages. The multilingual portion comes from a MarianMT translation pipeline. Non-Latin scripts such as Chinese, Japanese and Arabic are untested.
• Text-only. Image-conditioned queries are out of scope.
• Bounded by the teacher. Nothing in the objective lets the student exceed Qwen3-VL-Embedding-8B on the side it replaces.

Citation

@article{distilvdr2026,
  title   = {DistilVDR: A Compact End-to-End Visual Document Retriever
             via Dual-Student Distillation},
  author  = {Liu, Zhuchenyang and Wang, Ziyi and Zhang, Yao and Xiao, Yu},
  journal = {arXiv preprint arXiv:2608.10636},
  year    = {2026}
}

License

MIT.