decider-2b-vision: typed decisions from an image in one forward pass

The vision variant of decider-2b (v5 language weights transplanted into the full Qwen3.5-2B vision-language model), fine-tuned so that an image (a photo, a diagram, a game frame) plus a text question with lettered options yields a calibrated probability over the options at a single answer slot. No generation. A 256x240 game frame costs 64 visual tokens. Text-only questions work too, with decider-2b v5's behaviour, including its abstention handling.

Usage

import torch
from decider.vision import VisionDecisionModel
from decider.infer import Example, Q
m = VisionDecisionModel("<this repo>", grad_ckpt=False).cuda().eval()
ex = Example("This is a visual question about the image.",
             [Q("What is the person holding?", ["a phone", "a cup", "a book", "nothing"], 0)])
inp = m.prepare([(image, ex)])                  # image: PIL image, numpy array, or PNG bytes; None for text-only
probs = torch.softmax(m.slot_logits(inp), -1)[0, :4]

Training

One epoch (80k examples, 50k with images): game frames from Pong, Breakout, CliffWalking, MiniGrid and Super Mario Bros labelled by scripted policies (rare actions oversampled, plus DAgger frames from an earlier model's own play); multiple-choice image tasks from The Cauldron (A-OKVQA, AI2D, ScienceQA, IconQA, TQA, Raven, Hateful Memes); a replay of the text mixture. Then PPO from pixels on Breakout and Pong (the softmax over action options is the policy). Code: https://github.com/Mapika/decider

Results (300 items per task)

task	accuracy	ECE
Pong frames (agreement with the RAM-state teacher)	0.96	0.02
Breakout frames	0.96	0.02
Visual7W (held out)	0.89	0.03
A-OKVQA / AI2D / ScienceQA / IconQA / Raven / Hateful Memes	0.85 / 0.93 / 0.95 / 0.94 / 0.80 / 0.80	0.02 to 0.07

Playing from pixels only (no text state), three episodes each: Breakout 41 (the RAM-state teacher scores 22), Pong 3 (teacher 8), CliffWalking -13 (optimal), MiniGrid Empty 0.96 (teacher level); held-out Freeway 0, FrozenLake 0, the harder grid worlds 0 (their scripted teachers also score 0), Mario 1-1 315 px. The previous vision release (v4-based) scored Breakout 16, Pong 8, Freeway 8, BabyAI-GoTo 0.30; this one trades Pong and Freeway for Breakout and for the corrected abstention behaviour. Not intended as a chat model.