thaitea/laya-vision

Laya Vision

Typed, calibrated decisions about an image plus optional text, in one forward pass with no text generation. Give it a picture, some context and a set of questions. It answers each one as a multiple choice (choice), a yes/no probability (noul) or a level on a rubric you write (score), with probabilities you can act on at face value. It also plays simple games from pixels: the screen is the image and the buttons are the options.

This checkpoint is 201M parameters: SmolVLM-256M-Instruct cut to 20 of its 30 language layers, then trained for 2 hours on question-answering data and game frames. It was found by the autoresearch loop in the code repository.

It is the same model as thaitea/laya-vision-201m. The previous thaitea/laya-vision (237M, no game training) is still in this repo's history at revision d1fbdc0612fbe3b3d8ec6f54d328b195d35bb338 and in thaitea/laya-vision-smolvlm-256m-score.

• Code: github.com/r33drichards/laya-vision
• Community: Discord
• Status: experimental research. An independent fork of Laya, not affiliated with Convai Innovations, its authors.

Usage

git clone https://github.com/r33drichards/laya-vision && pip install -e ./laya-vision torchvision

import laya
from PIL import Image

agent = laya.load_vlm("thaitea/laya-vision")
result = agent.predict(
    {"image": Image.open("photo.jpg"), "note": "customer says it arrived broken"},
    {
        "damage":   {"type": "score",  "instructions": "How much damage does the item show?",
                     "criteria": ["none", "cosmetic: scratches or dents", "functional: parts broken or missing", "destroyed"]},
        "category": {"type": "choice", "instructions": "What kind of item is this?",
                     "criteria": ["electronics", "clothing", "furniture", "food", "other"]},
        "outdoors": {"type": "noul",   "instructions": "Was the photo taken outdoors?"},
    },
)
a = result["answers"]
a["damage"]["score"], a["category"]["choice"], a["outdoors"]["noul"]   # expected level 0-3, top option, P(true)

Pin a revision (laya.load_vlm("thaitea/laya-vision", revision=...)) if you need the weights not to change under you.

Results

Official validation splits

Calibrated answers, as predict returns them. These three numbers are backed by committed per-question predictions in the code repository (results/raw/), which benchmarks/verify_published.py recomputes.

Dataset	Question type	n	Accuracy	Previous checkpoint
A-OKVQA	4-way choice	1,138	59.8%	60.0%
ScienceQA, image subset	2–5-way choice	2,097	82.4%	82.8%
VQAv2 yes/no*	noul	5,000	71.4%	72.4%

* A re-split of the official VQAv2 validation set by image, the only official split with answers, so not comparable to published VQAv2 numbers.

Full evaluation suite

From the repository's full_eval run on this checkpoint (scorecard), next to the same suite on the previous checkpoint:

	This model	Previous checkpoint
Parameters	201M	237M
Accuracy, 59,427 questions over 34 validation sets	69.1%	69.3%
Mean accuracy over the 34 sets	71.6%	71.9%
Calibration error (ECE), all questions	0.041	0.064
Latency on an L4, same GPU, relative	0.83×	1.00×

On 23 of the 34 sets it is within a point of the previous checkpoint; it is ahead by more than a point on 3 and behind on 8. The largest drops are on the smallest sets: VQA-RAD (62 questions) 80.6% against 88.7%, InterGPS (94) 28.7% against 35.1%.

Games

Played greedily from pixels, one forward pass per move, on seeds never used in training. The score is normalized per game: 0 is random play, 1 is the scripted or trained expert (the autoresearch benchmark, clipped to [−0.5, 1.5]):

Game	Normalized score
ViZDoom basic	0.99
Atari Freeway	0.81
Acrobot	0.76
MountainCar	0.64
Maze 4×4	0.32
Atari Breakout	0.20
Snake 10×10	0.17
Maze 6×6	0.02
CartPole	−0.01
LunarLander	−0.44
Mean	0.35

The previous checkpoint was never trained on games and scores −0.04 on the same benchmark.

Training

• Start: the previous thaitea/laya-vision checkpoint, cut to its first 20 language-model layers. The vision tower (12 layers) is kept and frozen.
• Data: every train record of 19 closed-answer subsets of The Cauldron and 4 rubric-scored sets (VLFeedback, AVA, RichHF-18K, CrisisMMD), minus a calibration tail. 45% of training draws were game frames:
  • Maze, Snake and classic control (CartPole, Acrobot, MountainCar, LunarLander), generated with every best move as a soft target;
  • Atari Freeway and Breakout (a PPO expert's action distributions) and ViZDoom basic (a scripted expert).
• Objective: Laya's: soft cross-entropy plus a strictly proper scoring rule, with the option order shuffled at random.
• Schedule: 28,414 steps at batch 64 on one H100, 120 minutes (1.82M samples). Head LR 5e-5, backbone LR 1e-5, 20 warmup steps, cosine decay to 10%.
• Calibration: per-type temperatures fitted on the held-out calibration tail: choice 3.86, score 2.10, noul 3.05.
• How the recipe was chosen: the autoresearch loop trains candidate recipes for 15 minutes and keeps the ones that extend a Pareto frontier over quality, games, parameters and latency. The 2-hour run is its batch-64 frontier point trained 8× longer.

Limitations

• Games are uneven. LunarLander got worse with training, and 6×6 mazes are rarely solved.
• Small sets dropped. VQA-RAD and InterGPS are several points below the previous checkpoint (few questions, so noisy, but real).
• Human vote spreads. On sets scored against how people split their votes, it beats a predict-the-average baseline on only 3 of 9.
• Domain. Trained on everyday photos, diagrams, charts, rubric-graded images and simple game screens. Calibrate on your own labelled data (agent.calibrate) before acting on probability thresholds in another domain.
• One image per state, resized to a single 512 px tile.

License

The weights are released under CC BY-NC-SA 4.0, because the training data includes ScienceQA and CrisisMMD, which carry that license. The base model, SmolVLM, is Apache 2.0. The code is Apache 2.0.