Sieve-Vision-4B

Sieve-Vision-4B is the first vision model of the Sieve family. You give it one or more images and a multiple-choice question, and it returns a calibrated probability for every option. It reads the images and the question once, scores the options directly, and never generates text.

It is a full fine-tune of Qwen/Qwen3.5-4B-Base (revision 1001bb4d) with a small decision head. The inference code is in this repository (sieve_vision/), including an engine for the Decision Index kit.

Summary

How it works

• Backbone. Qwen3.5-4B-Base: a 24-layer vision encoder (16-px patches, 2×2 merge, so one token per 32×32 px) feeding a 32-layer text model (24 Gated DeltaNet layers and 8 full-attention layers).
• Prompt. Image k: <image tokens> for every image, then Question: and the question, then Options: with one line per option, - <key>: <option text>, then Answer:.
• Readout. The decision head reads the hidden state at Answer: and the mean hidden state of each option's own tokens. It scores each option with a bilinear term plus the backbone's own next-token logit for the option's key, scaled by a learned weight (0.985 after training). A logit cap and a calibrated softmax give the probabilities.
• One pass. One forward pass per question; nothing is generated, and the options are scored in the order given.

Use

pip install "torch>=2.8" "transformers>=5.16" safetensors pillow huggingface_hub
pip install flash-linear-attention   # recommended: fused kernels for the Gated DeltaNet layers (our runs used 0.5.2)

import sys
from huggingface_hub import snapshot_download

path = snapshot_download("sthanika-ai/Sieve-Vision-4B")
sys.path.insert(0, path)
from sieve_vision import SieveVisionEngine

engine = SieveVisionEngine(model=path)          # or model="sthanika-ai/Sieve-Vision-4B"
response, _ = engine(
    {},                                          # state (optional text or JSON shown before the question)
    {"answer": {"type": "choice", "instructions": "Is the parcel in the photo damaged?",
                "criteria": {"A": "Yes, the box is torn or crushed", "B": "No, it looks intact"}}},
    images=["parcel.jpg"])
print(response["answers"]["answer"])  # {'type': 'choice', 'choice': 'A', 'probabilities': {'A': ..., 'B': ...}}

• Images. Paths, data: URLs, raw base64 or bytes. Images above 1.6 MP are downscaled (LANCZOS, aspect kept); smaller ones are used at their own size.

• Batching. engine.batch([{"state": ..., "questions": ..., "images": [...]}, ...]) answers many requests at once.

• Decision Index.

  hf download sthanika-ai/Sieve-Vision-4B --local-dir Sieve-Vision-4B
  PYTHONPATH=Sieve-Vision-4B decision-index vision pipeline --dir work/vision \
      --engine sieve_vision_engine:SieveVisionEngine --option model=sthanika-ai/Sieve-Vision-4B \
      --out runs/Sieve-Vision-4B
  

Decision Index (vision)

Our own run of the complete public vision suite of the Decision Index 0.3 (12,210 requests, 11 benchmarks) with the official kit at f5a6e4f; the suite was rebuilt from the pinned sources and passes decision-index vision verify. The kit's scorer gives a chance-corrected skill per benchmark (0 = random guessing, 100 = perfect) and the board's weights (R-Bench-M 0.5, MMMU-Pro vision 0.25, the rest 1). Files: sthanika-ai/Sieve-Vision-4B-decision-index-results.

All 12,210 requests were answered: 0 unsupported, 0 errors. The leaderboard's Full score adds private tests of the same skills, which the maintainers run themselves.

Training

Training data

114,822 training and 5,795 dev multiple-choice questions over 1 to 10 images, all from public datasets whose licences allow commercial use. Questions, answers and images come from the sources; the only conversion is into the multiple-choice format (yes/no items become two options; Mind2Web uses the board's own set-of-marks conversion on the train split; TextVQA and PlotQA distractors are other text or answers from the same image).

Decontamination. Every training and dev image was compared with every image of the built public vision suite (exact SHA-256, and near-duplicate when pHash ≤ 4 and dHash ≤ 10), and question texts with the suite's question texts (13-gram overlap). The final check over all 119,000+ training and dev images flagged 11 near-duplicates and no exact match: 9 are blank or single-colour images, and 2 are real photos that appear in unrelated training questions (one Hateful Memes image, one Winoground image). Rescoring without the affected suite rows gives 64.94. The report is in the results dataset (reports/contamination_check.json).

Evaluation on our own held-out data

Limitations

• Hateful content. No licence-compatible meme moderation data exists, so the model was not trained for it; its Hateful Memes skill is low (22.2). Do not use it for content moderation decisions.
• Expert reasoning. College- and expert-level questions (R-Bench-M, MMMU-Pro) remain weak.
• Choice only. It answers multiple-choice questions over images; it does not generate text or captions.
• English. Trained on English questions only.
• Third-party images. Several training sources use images whose copyright belongs to third parties (COCO, Visual Genome, web screenshots), under their datasets' terms.

Licence

Apache-2.0, as the backbone (Qwen/Qwen3.5-4B-Base); the LICENSE file is the backbone's. Training data licences are listed above.

Citation

@misc{sieve_vision_2026,
  title  = {Sieve-Vision-4B: a one-pass vision decision model},
  author = {Sthanika AI},
  year   = {2026},
  url    = {https://huggingface.co/sthanika-ai/Sieve-Vision-4B}
}