thaitea/laya-vision-modernvbert-250m

Laya Vision (ModernVBERT-250M)

This model makes calibrated, typed decisions about one or more images plus optional text. It answers choice, score and noul (yes/no probability) questions in one forward pass, with no text generation.

It adds image input to Laya by replacing Laya's ModernBERT text encoder with ModernVBERT, a bidirectional ModernBERT-150M plus SigLIP2 vision encoder. Because the backbone is bidirectional, Laya's original readout carries over unchanged: each option is scored from a [MASK] marker that sees the whole sequence. Laya's predict(state, questions) API, proper-scoring-rule training and temperature calibration are also unchanged. A sibling checkpoint on a causal backbone is thaitea/laya-vision-smolvlm-256m.

• Code: github.com/r33drichards/laya-vision, see the "ModernVBERT experiment" and "Post-training on The Cauldron" sections and docs/modernvbert-cauldron.md.
• Status: experimental. This is an independent research fork, not affiliated with Convai Innovations, the authors of Laya, nor with the ModernVBERT authors.

Usage

Needs transformers >= 5.3 for the ModernVBERT classes.

git clone https://github.com/r33drichards/laya-vision && pip install -e ./laya-vision torchvision "transformers>=5.3"

import laya
from PIL import Image

agent = laya.load_vlm("thaitea/laya-vision-modernvbert-250m")
result = agent.predict(
    {"image": Image.open("photo.jpg"), "note": "customer says it arrived broken"},
    {
        "damaged":  {"type": "noul",   "instructions": "Does the item in the photo look damaged?"},
        "category": {"type": "choice", "instructions": "What kind of item is this?",
                     "criteria": ["electronics", "clothing", "furniture", "food", "other"]},
    },
)
result["answers"]["damaged"]["noul"]        # calibrated P(true)
result["answers"]["category"]["choice"]     # top option; see ["probabilities"], ["confidence"]

A state may carry several images: {"images": [left, right]} (the model was trained on NLVR2 pairs).

Results

Scores are on full validation splits. "Calibrated" uses the per-type temperatures stored in vlm_agent_config.json, which are applied automatically. The first three rows are the official splits the SmolVLM checkpoint reports on; the model saw those datasets' train rows only as 3 of its 19 training subsets.

Dataset	Question type	Chance	Accuracy	ECE raw	ECE calibrated
A-OKVQA (n=1,138)	4-way choice	25%	65.2%	0.214	0.064
ScienceQA, image subset (n=2,097)	2–5-way choice	~36%	79.0%	0.114	0.058
VQAv2 yes/no (n=5,000)*	noul	50%	71.8%	0.134	0.037
All 22 validation sets (n=22,886), the three above plus a 5% holdout of each training subset			71.8%	0.103	0.022

* The VQAv2 train/val split is a re-split of the official VQAv2 validation set by image (the only official split with answers), so these numbers are not comparable to published VQAv2 results.

Per-subset holdout accuracy ranges from 93% (DVQA) and 86% (OCR-VQA) down to 40% (RAVEN, 8-way) and 27% (InterGPS geometry); the full table is in training_metrics.json and in the repo's docs/modernvbert-cauldron.md. Against the SmolVLM checkpoint, which was trained for 3 epochs on exactly the three official sets, this model is 3.4 points better on A-OKVQA and 7.6 and 1.6 points worse on ScienceQA and VQAv2 yes/no.

• Latency: about 32 ms for one image question on an NVIDIA L4 (bf16), 48 ms in fp32. The image is encoded once per predict call and shared by every question.
• Option order: across 4 rotations of the A-OKVQA option order, accuracy varies by 1.0 point. Averaging over orders (n_permutations=4) gives 66.4%.

Training

• Backbone: ModernVBERT (ModernBERT-150M text encoder, SigLIP2-100M vision tower, pixel-shuffle connector, 512-pixel tiles, 64 image tokens). The vision tower is frozen; the text encoder, connector and decision head are trained.
• How options are read: [CLS]User:<image tokens> <type> question: <instructions>[SEP][MASK] option 1[MASK] option 2 ...[SEP]<state>[SEP]. Each option is scored from its [MASK], as in Laya's text model; the act head pools [CLS].
• Loss: soft cross-entropy plus a strictly proper scoring rule (log + spherical), option order shuffled at random.
• Data: the closed-form turns of 19 subsets of The Cauldron, 269,900 questions after holdouts: lettered multiple choice (AI2D, IconQA, InterGPS, ScienceQA, TQA, Visual7W), A-OKVQA's options lists, RAVEN's lettered panels, and yes/no turns (FigureQA, Hateful Memes, NLVR2, VSR, VQA-RAD, and the yes/no share of CLEVR, DVQA, MapQA, OCR-VQA, VQAv2, ChartQA). Up to 10,000 rows per subset, at most 4 questions per row. Subsets were sampled equally and none was repeated more than 4 times.
• Schedule: 16,869 steps at batch 32 on one A100 (68 minutes). Head LR 1.4e-4, backbone LR 2.8e-5, 3% warmup, then cosine decay to 10%. Two epochs; the final checkpoint had the best mean validation accuracy.
• Calibration: per-type temperatures fitted on the last 300 training records of each subset, held out of training: choice 2.71, noul 2.16.
• training_metrics.json has every evaluation from the run.

Limitations

• score questions are untrained. There was no ordinal image data, so their outputs are meaningless.
• Overconfident before calibration. Raw ECE on the choice sets is 0.11 to 0.21; rely on the calibrated outputs, which the loader applies by default.
• Weak on visual reasoning. Geometry (InterGPS 27%), abstract analogies (RAVEN 40%) and textbook diagrams (TQA 51%) are far below the chart, document and photo sets.
• Images are 512-pixel tiles. Fine print in documents may be lost; the image processor's tiling is turned off.
• Text-only states work but are unevaluated. This model is not a drop-in replacement for Laya's text checkpoint.

License

The weights are released under CC BY-NC-SA 4.0, because the training data includes ScienceQA, which uses that license. The base model, ModernVBERT, is MIT. The Cauldron's subsets each carry their own licenses; consult the dataset card before commercial use of any derivative. The code is Apache 2.0.