DNA-VL-STEER-2B

STEER — Steering Trajectories onto the Established Residual-stream.

DNA-VL-STEER-2B is a language-bias–calibrated variant of Qwen/Qwen3-VL-Embedding-2B. It keeps everything that makes the base model good — the same architecture, the same 2048-dim shared image/text/video embedding space, the same instruction-aware interface — and closes most of the cross-lingual retrieval gap: low-resource languages that lagged far behind English in the frozen model are pulled up sharply, while English, the other high-resource languages, and the vision side are preserved.

What's new in this release. This version is calibrated on all 36 XM3600 languages (previous releases used 11). Broadening the calibration set does two things: it lifts the low-resource languages further, and — crucially — it extends the gains to the languages a language-limited calibration left behind. On the full 36-language XM3600 pool, mean image→text R@1 goes 59.8 → 70.3 and text→image 48.7 → 59.5. The 25 languages that were outside the earlier 11-language calibration (and actually regressed under it) now improve from 61.4 → 72.2 (i2t).

It is a drop-in replacement: load it exactly like the base model and the API is identical. Under the hood only the early residual-stream layers are adjusted; every other weight (including the entire vision tower) is bit-identical to the base.

• Base model: Qwen/Qwen3-VL-Embedding-2B (28 layers, 2048-dim, 32K context, MRL 64–2048)
• Modalities: text · image · screenshot · video · arbitrary interleaved combinations
• Calibrated languages (36): en, fr, de, es, zh, ko, bn, fil, hi, sw, te, ja, ru, ar, vi, tr, th, el, fa, fi, id, he, pl, uk, nl, it, pt, cs, sv, hu, ro, hr, da, no, mi, quz — i.e. the full XM3600 language set, spanning high-resource, low-resource, and previously-unseen languages.
• Korean improves across every retrieval benchmark (+4 to +7 R@1), image and video alike — see the dedicated section below.

---

Usage

Because the release copies the base model's Sentence-Transformers configuration verbatim, usage is identical to Qwen/Qwen3-VL-Embedding-2B — only the model id changes.

Requires transformers >= 5.0 — older versions load RANDOM weights silently

On transformers 4.57.x and earlier, Qwen3VLModel.base_model_prefix is "" while the checkpoint stores its 625 tensors under a model. prefix, so zero keys match and the model is randomly initialized. There is no error — only a newly initialized warning that is easy to miss, and embeddings that look plausible but are meaningless. This affects the base Qwen/Qwen3-VL-Embedding-2B identically; it is not specific to STEER.

pip install "transformers>=5.0" sentence-transformers qwen-vl-utils

Verify your install in one line — if this does not print True, your weights are random:

import torch
from sentence_transformers import SentenceTransformer
m = SentenceTransformer("dnotitia/DNA-VL-STEER-2B")
a = m.encode(["a photo of a cat"]); b = m.encode(["a picture of a cat"])
c = m.encode(["quarterly earnings fell"])
print(float(a @ b.T) > float(a @ c.T) + 0.1)   # must be True

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("dnotitia/DNA-VL-STEER-2B")

1. Text alone

sentences = [
    "고양이가 창턱에서 자고 있다.",           # Korean: "A cat is sleeping on the windowsill."
    "Un chat dort sur le rebord de la fenêtre.",  # French, same meaning
    "The stock market fell sharply today.",
]
emb = model.encode(sentences)          # (3, 2048), L2-normalized
print(model.similarity(emb, emb))      # the KO/FR pair scores high; the finance line low

2. Image alone

images = [
    "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
    "/path/to/local/photo.jpg",        # local paths work too
]
img_emb = model.encode(images)         # (2, 2048)

3. Interleaved (text + image in one input)

Pass a dict with text and/or image. This is how you embed a captioned image, a screenshot with a question, a document page with a title, etc.

documents = [
    {"text": "A golden retriever offering its paw on a beach at sunset.",
     "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"},
    {"text": "City skyline at night."},                       # text-only entry
    {"image": "/path/to/local/photo.jpg"},                    # image-only entry
]
doc_emb = model.encode(documents)      # (3, 2048)

4. Cross-lingual, cross-modal retrieval

All modalities and languages share one space, so a query in any language retrieves matching images/text directly:

queries = [
    "해변에서 강아지와 노는 여자",          # Korean query
    "mujer jugando con su perro en la playa",  # Spanish query
]
gallery = [
    "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
    {"text": "A woman plays with her golden retriever on a sunlit beach."},
    {"text": "A city skyline viewed from a rooftop at night."},
]
q = model.encode(queries)
g = model.encode(gallery)
print(model.similarity(q, g))          # both queries rank the beach image/caption first

5. Video

Pass a video key. An .mp4 path works and a list of frame paths works — a directory does not. The two working routes are not the same protocol, so pick deliberately:

you pass	sampling	matches our reported numbers?
"/path/to/clip.mp4"	fps=1, capped at 64 frames	no — close, but not identical
["f0.jpg", "f1.jpg", ...]	uniform 64 + pixel budget	yes — this is what we benchmarked
"/path/to/frames_dir"	—	no — fails silently, see warning

# convenient: let the loader decode the file (fps=1, capped at 64 frames)
vid_emb = model.encode([{"video": "/path/to/clip.mp4"}])

# reproducible: uniform 64 frames, the protocol behind every video number in this card
import decord, numpy as np
vr  = decord.VideoReader("/path/to/clip.mp4")
idx = np.linspace(0, len(vr) - 1, 64).astype(int)
frames = [...]                                   # save/collect those 64 frames as paths or PIL
vid_emb = model.encode([{"video": frames}])

# text -> video retrieval: same instruction on both sides
q = model.encode(["a person kneading dough on a floured counter"])
print(model.similarity(q, vid_emb))

Never pass a directory path as video

The video backends cannot open a directory, the error is swallowed, and you get an embedding that does not depend on the input at all — we measured cosine 1.0000 between three different clips passed this way (the same clips as frame lists: 0.57–0.67). Nothing raises, so this corrupts a retrieval index silently. Expand the directory to a sorted list of frame paths yourself: {"video": sorted(glob.glob(f"{d}/*.jpg"))}.

Uniform 64 frames is the protocol behind the reported video numbers; fewer will cost you accuracy on anything temporal. Note the .mp4 route samples at 1 fps, so a clip shorter than 64 s yields fewer than 64 frames — for short clips the two routes diverge most.

6. Zero-shot classification

Put the task instruction on the media side and leave the class names on the neutral default. Do not wrap labels in a template, and do not lowercase them — see the measured effects below.

labels = ["ApplyEyeMakeup", "Knitting", "PlayingGuitar"]   # keep original casing
lab_emb = model.encode(labels)                             # neutral default prompt

vid_emb = model.encode(
    [{"video": "/path/to/clip.mp4"}],
    prompt="Recognize the category of the video content.",  # task instruction here
)
pred = labels[int(model.similarity(vid_emb, lab_emb).argmax())]

7. Task instructions & custom dimensions (inherited from the base)

The model is instruction-aware and supports Matryoshka (MRL) output dimensions (64–2048). Per Qwen's guidance, write instructions in English even for non-English inputs.

# custom instruction — see the measured effect below
q = model.encode(queries, prompt="Retrieve the image that matches this description.")

# shorter embeddings via MRL — truncate + renormalize, no second model load
import numpy as np
emb512 = emb[:, :512]
emb512 /= np.linalg.norm(emb512, axis=1, keepdims=True)

# (SentenceTransformer(..., truncate_dim=512) also works, but loads the model a second time —
#  slicing an existing embedding is identical and free.)

How much does the instruction matter? Measured on video classification, swapping the neutral default for a task instruction ("Recognize the category of the video content."):

task	labels	gain
Kinetics-700	519 fine-grained	+3.8
SomethingSomethingV2	175, fragments like holding pen	+3.7
Breakfast	10 bare nouns	+1.8
UCF101	101 distinctive compounds	+0.1
HMDB51	51	0.0

The gain scales with how ambiguous the label set is — large or fragmentary label sets benefit, already-distinctive labels do not. A CLIP-style template ("a video of {}") does not help and slightly hurts once a proper instruction is set.

Keep label capitalization. On UCF101, lowercasing class names costs −5.4 and splitting CamelCase costs a further −1.5; ApplyEyeMakeup (94.3) beats apply eye makeup (91.4) beats applyeyemakeup (88.9). Capitalization does word-segmentation work the tokenizer relies on.

Pro tip — sharp reproduction

STEER inherits the base model's exact preprocessing protocol, and the reported numbers depend on it. To reproduce cleanly:

• Load in bfloat16 and L2-normalize outputs (Sentence-Transformers does this by default).
• Keep the official image pixel budget — min_pixels = 4·32², max_pixels = 1800·32², patch size 16, and do not let your image loader pre-downscale; feeding smaller images than the protocol is the most common cause of "my numbers are lower than the card".
• Pooling is last non-pad token of the post-final-RMSNorm hidden state — don't substitute mean-pooling.
• Check your max_length against your actual documents. Harnesses commonly default to 8192 (the official MMEB value) while this model has a 32K context. On a long-document benchmark that silently truncates: we measured 13.5 against a published 94.4 purely because documents ran to ~24.5k words and the answer sat past the cut.
• Fix the batch size when two runs must match exactly. Right padding plus last-token pooling makes embeddings mildly batch-composition dependent (~5e-3), which is ±0.2–0.7 on a small benchmark. Same batch size ⇒ bit-identical results.
• Instructions in English, even for non-English queries, and reuse the same instruction across the query/document sides you intend to compare.
• The calibration boost covers the 36 languages listed above — expect the largest lifts on low-resource languages and preservation on English and the vision side (see below).

---

Performance

All numbers are frozen base → STEER on the official post-norm protocol. STEER's design goal is to lift low-resource languages without paying for it on English or vision, so read the tables as: big LRL gains, high-resource preserved, small honest costs where they exist.

Korean capability

Korean is one of the calibrated languages, and it gains sharply in both directions on every retrieval benchmark — image and video — while nothing about the English side is sacrificed. If your workload is Korean-centric, STEER is a clear upgrade over the base model.

Korean image↔text retrieval (R@1, frozen → STEER):

Benchmark	image→text	text→image
XM3600 (full pool)	70.2 → 75.0 (+4.8)	57.9 → 64.2 (+6.3)
Flickr30k-1k	71.3 → 78.7 (+7.4)	68.4 → 79.5 (+11.1)
XTD-1k	58.8 → 64.9 (+6.1)	53.6 → 63.8 (+10.2)

Korean zero-shot video retrieval (R@1, frozen → STEER) — trained only on image–text, transfers to video:

Dataset	text→video	video→text
MSR-VTT	39.7 → 43.6	39.7 → 42.4
DiDeMo	38.4 → 43.1	34.9 → 43.0
MSVD	47.4 → 53.7	73.6 → 78.2
VATEX	31.5 → 38.3	47.7 → 55.4

Multilingual image↔text retrieval — XM3600 (R@1, full 3,600-image pool)

Calibrated on all 36 languages, the gains now cover the entire XM3600 language set — the low-resource tier jumps, and the languages that were previously outside the calibration (and lost ground under a language-limited calibration) now improve too.

Tier	image→text	text→image
HRL-6 (high-resource)	75.3 → 77.8	64.7 → 67.4
LRL-5 (low-resource)	32.8 → 51.8 (+19.0)	21.2 → 41.0 (+19.8)
+25 additional languages	61.4 → 72.2 (+10.8)	50.4 → 61.2 (+10.8)
Overall-36	59.8 → 70.3 (+10.5)	48.7 → 59.5 (+10.8)

_{Protocol: every caption is used as a query (~7,200 per language) against the full 3,600-image pool. A first-caption-only protocol gives lower i→t absolutes because there is then one relevant caption per image instead of ~2; the frozen→calibrated deltas are the same either way. These figures are computed by the same harness as the benchmark report attached to this repo, so the two agree cell for cell.}

The high-resource↔low-resource R@1 gap (i→t) narrows from 42.6 → 25.9.

Cross-benchmark multilingual retrieval (R@1)

Benchmark	HRL i→t	LRL i→t	HRL−LRL gap
Flickr30k-1k	79.1 → 82.9	48.7 → 74.5 (+25.8)	30.4 → 8.4
XTD-1k	67.5 → 71.4	42.5 → 63.3 (+20.8)	25.1 → 8.1

Zero-shot cross-modal transfer — video retrieval (Text→Video R@1)

Trained only on image–text, yet low-resource video retrieval improves and English is preserved.

Dataset	en	ko	zh
MSR-VTT	53.1 → 50.7	39.7 → 43.6	47.6 → 49.6
DiDeMo	54.8 → 53.9	38.4 → 43.1	48.5 → 49.4
MSVD	61.1 → 60.8	47.4 → 53.7	53.6 → 56.8
VATEX	47.8 → 47.5	31.5 → 38.3	44.2 → 44.6

MIEB (Multilingual) — 130-task battery, Overall = mean of 10 category means

	Overall	Multilingual retrieval	(other 9 categories)
frozen	66.71	64.80	preserved
STEER	67.86	71.98	preserved

The multilingual-retrieval category rises +7.18 while the nine non-target categories move within noise: 9 of 10 categories improve and the only decline is document understanding at −0.33. Recomputed on the identical 130 tasks for every model with full coverage.

_{Overall is the mean of the 10 task-type means, per the MIEB(Multilingual) definition — not a flat mean over the 130 tasks (a flat mean gives 60.06 → 60.53 and understates the effect, because the 3-task multilingual-retrieval category is swamped by the 45-task general-retrieval one). English visual STS is 79.88 → 80.24; an earlier version of this card reported 77.2 → 80.2 for that row, which came from a mis-slice of the VisualSTS aggregate tasks.}

Document retrieval (preservation check)

Benchmark	frozen → STEER
ViDoRe (v3) (nDCG)	52.9 → 53.2

Multilingual visual-document retrieval is unchanged — calibration does not disturb the document-understanding capability of the base model.

Honest limitations

• Calibration targets the multilingual visual-caption domain. The gains are largest on multilingual image/text/video retrieval across the 36 calibrated languages. Languages entirely outside this set, and abstract text-only domains far from visual captions, should be expected to see little change rather than a large lift.

• Breadth trades against per-language peak. Spreading calibration across all 36 languages gives each low-resource language a slightly smaller individual boost than a calibration narrowly focused on a handful of them would, but coverage is uniform across the language set.

• Some tasks regress — pick the variant per workload. An earlier version of this card said nothing regresses relative to the frozen base. That is not correct, and broader evaluation since has measured where it does:

  workload	STEER vs frozen
  Multilingual retrieval (XM3600, ko/zh video)	+5 to +11
  Abstract text retrieval (Wikipedia multilingual)	−7.9 R@1 (2B), −4.1 (8B)
  Video action classification	−4 to −9
  Fine-grained visual (logo retrieval)	−3.1 mAP

  The calibration is visual-vocabulary specific: it helps caption-like multilingual retrieval and costs abstract text and fine-grained visual discrimination. If your workload is English, text-only, or fine-grained classification, use the frozen base model. The effect also attenuates with size — at 8B the gains shrink and several tasks turn net-negative.

• Composed queries are not supported. "This video, but with X changed" does not work: on Dense-WebVid-CoVR both an instruction-conditioned composition and a vector-sum composition scored below using the edit target's caption alone. The model was trained for symmetric similarity, not edit-application; no inference-time prompt recovers it.

• Matryoshka truncation is task-dependent. Degradation is smooth and monotonic at every width, but how much you can cut varies enormously: text retrieval keeps 99.6% of its score at 64 dims, while fine-grained logo retrieval keeps only 49.5%. At 512 dims the median cost across 15 benchmarks is 1.6%. Measure on your own task before truncating hard.

---

Derived from Qwen/Qwen3-VL-Embedding-2B. Vision tower and all layers outside the early residual-stream are unchanged from the base model. Calibrated on the full 36-language XM3600 set.