SuperGlance-9B-Stage2

Paper: arXiv:2611.19138 ยท PDF ยท Project page: guobapei.github.io/SuperGlance ยท Code: GuobaPei/SuperGlance

The release terms of this repository are not decided yet; see Licence below.

This is SuperGlance's stage-2 model and the start model of its driving stages (supervised training, its last phase and RL, which give SuperGlance-9B). It is Qwen/Qwen3.5-9B with two LoRA stages merged in. Each merge is Wโ€ฒ = round_bf16(W + 2.0 ยท BยทA), and every other tensor is unchanged. It loads like the base model, with AutoModelForImageTextToText.

  • Stage 1 drew 500,000 questions from a pool of 1,158,150 multiple-choice questions. The sources are:
    • SAT, GQA, TallyQA, ChartQA, AI2D and TextVQA;
    • OpenImages box choice and SpatialSense;
    • 13 text QA datasets.
  • Stage 2 drew 192,000 questions from a pool of 1,240,040: the stage-1 questions except text_nota and openimages_vertical (1,115,662), plus 124,378 DriveLM-nuScenes questions.

In both stages, each drawn question is shown in its stored option order and a shifted one.

The questions and their format are in the SuperGlance-QA dataset. The training code is in the SuperGlance code release.

Files

file content
model.safetensors-*-of-00004.safetensors, model.safetensors.index.json, config.json, tokenizer and processor files the merged model (bf16)
LICENSE-Qwen3.5-9B the base model's licence (Apache-2.0), kept as its terms require

Licence

The release terms are pending:

  • The base model, Qwen3.5-9B, is Apache-2.0.
  • The training data's terms are listed per source in SuperGlance-QA. They include DriveLM-nuScenes, which is CC BY-NC-SA 4.0 plus the nuScenes terms of use.
Downloads last month
63
Safetensors
Model size
10B params
Tensor type
BF16
ยท
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for pei2333/SuperGlance-9B-Stage2

Finetuned
Qwen/Qwen3.5-9B
Finetuned
(1)
this model
Finetunes
1 model

Paper for pei2333/SuperGlance-9B-Stage2