Contamination MapText / LLM / Model
Qwen3-VL
- Organisation
- Alibaba Cloud (Qwen Team)
- Family
- Qwen · version 3
- Released
- 2025-09-22 (HF repo creation of Qwen/Qwen3-VL-235B-A22B-Instruct (2025-09-22, via huggingface.co API); smaller sizes followed until 2025-10-19; tech report arXiv v1 2025-11-26)
- Checkpoints
- Qwen/Qwen3-VL-2B-Instruct, Qwen/Qwen3-VL-2B-Thinking, Qwen/Qwen3-VL-4B-Instruct, Qwen/Qwen3-VL-4B-Thinking, Qwen/Qwen3-VL-8B-Instruct, Qwen/Qwen3-VL-8B-Thinking, Qwen/Qwen3-VL-32B-Instruct, Qwen/Qwen3-VL-32B-Thinking, Qwen/Qwen3-VL-30B-A3B-Instruct, Qwen/Qwen3-VL-30B-A3B-Thinking, Qwen/Qwen3-VL-235B-A22B-Instruct, Qwen/Qwen3-VL-235B-A22B-Thinking
- Weights
- Open · Apache 2.0
- Links
- PaperModel cardCode
Training data disclosure: Vague
Level determined for 0% of benchmark cells.
- TestTest data seen0
- SeenSeen, split unclear0
- TrainTraining data seen0
- Related data seen0
- NoNot seen0
- ?Unknown22
“The model is trained on a massive and diverse dataset of approximately 1 trillion (1T) tokens.”
Built on
Everything these base models may have seen is inherited by Qwen3-VL.
-
Qwen3
· disclosure: Vague
Evidence
“all built upon Qwen3 backbones”
arxiv.org · paper · Section 2 Large Language Model · retrieved 2026-10-15 -
SigLIP 2
· disclosure: Partial list
SigLIP2-SO-400M by default; SigLIP2-Large (300M) for the 2B and 4B models.
Evidence
“We utilize the SigLIP-2 architecture (Tschannen et al., 2025) as our vision encoder and continue training it with dynamic input resolutions, initialized from official pretrained checkpoints.”
arxiv.org · paper · Section 2 Vision Encoder · retrieved 2026-10-15
Other declared training data
-
web image-caption pairs (predominantly Chinese-English, recaptioned) (broad web data: may contain benchmark test sets) · split: unknown
Evidence
“We curate a large-scale corpus of contemporary, predominantly Chinese–English multilingual image–text pairs from web sources”
arxiv.org · paper · Section 3.2.1 · retrieved 2026-10-15 -
interleaved image-text documents from Chinese and English websites (broad web data: may contain benchmark test sets) · split: unknown
Evidence
“We collect diverse real-world multimodal documents sourced from recent Chinese and English websites”
arxiv.org · paper · Section 3.2.1 · retrieved 2026-10-15 -
3 million PDFs from Common Crawl (document parsing) and 4 million internal documents (broad web data: may contain benchmark test sets) · split: unknown
Evidence
“For document parsing, we collect 3 million PDFs from Common Crawl, evenly distributed across 10 document types (300K samples each), along with 4 million internal documents.”
arxiv.org · paper · Section 3.2.3 · retrieved 2026-10-15 -
COCO, Objects365, OpenImages, RefCOCO/+/g (box grounding) (not yet matched to a dataset) · split: unknown
Evidence
“We begin by aggregating widely used open-source datasets, including COCO (Lin et al., 2014), Objects365 (Shao et al., 2019), OpenImages (Kuznetsova et al., 2020), and RefCOCO/+/g”
arxiv.org · paper · Section 3.2.4 · retrieved 2026-10-15 -
PixMo pointing and counting annotations (not yet matched to a dataset) · split: unknown
Evidence
“public pointing and counting annotations from PixMo (Deitke et al., 2024)”
arxiv.org · paper · Section 3.2.4 · retrieved 2026-10-15 -
K-12 and undergraduate-level exercises (multimodal reasoning data, >60M) (not yet matched to a dataset) · split: unknown
Sources of the exercises are not named.
Evidence
“million K–12 and undergraduate-level exercises, meticulously curated through a rigorous cleaning and reformulation pipeline”
arxiv.org · paper · Section 3.2.8 · retrieved 2026-10-15 -
synthesized long-CoT multimodal reasoning samples (>12M) (not yet matched to a dataset) · split: unknown
Evidence
“Regarding long CoT problem-solving data, we synthesize over 12 million multimodal reasoning samples paired with images.”
arxiv.org · paper · Section 3.2.8 · retrieved 2026-10-15 -
Qwen3 and Qwen3-Coder code corpus (not yet matched to a dataset) · split: unknown
Evidence
“We reuse the extensive code corpus from the Qwen3 and Qwen3-Coder series.”
arxiv.org · paper · Section 3.2.6 · retrieved 2026-10-15 -
SFT dataset (~1.2M samples from open-source datasets and web resources) (not yet matched to a dataset) · split: unknown · fine-tuning
Open-source datasets are not named.
Evidence
“we systematically curated the SFT dataset through the meticulous selection and synthesis of samples from open-source datasets and web resources.”
arxiv.org · paper · Section 4.2.1 · retrieved 2026-10-15 “This dataset comprises approximately 1,200,000 samples”
arxiv.org · paper · Section 4.2.1 · retrieved 2026-10-15 -
RL queries from open-source and proprietary sources (~30K) (not yet matched to a dataset) · split: unknown · fine-tuning
Evidence
“We curate training data from both open-source and proprietary sources and apply rigorous preprocessing and manual annotation to ensure high-quality RL queries.”
arxiv.org · paper · Section 4.4.1 · retrieved 2026-10-15
Benchmark cells
Classic LLM benchmarks
Unknown MMLU (Measuring Massive Multitask Language Understanding) · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown GSM8K (Grade School Math 8K) · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown MATH (Mathematics Aptitude Test of Heuristics) · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown HumanEval · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown ARC-Challenge (AI2 Reasoning Challenge, Challenge Set) · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown HellaSwag · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown TruthfulQA · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown WinoGrande · unknown for its base model Qwen3: training data not disclosed in enough detail
Multimodal LLM benchmarks
Unknown MMMU (Massive Multi-discipline Multimodal Understanding and Reasoning) · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown MMMU-Pro · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown MathVista · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown ChartQA · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown DocVQA · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown AI2D (AI2 Diagrams) · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown CharXiv · unknown for its base model Qwen3: training data not disclosed in enough detail
Recent LLM benchmarks
Unknown GPQA Diamond · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown MMLU-Pro · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown LiveCodeBench (code generation, release_v6) · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown SWE-bench Verified · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown AIME 2024 · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown AIME 2025 · unknown for its base model Qwen3: training data not disclosed in enough detail
Unknown MATH-500 · unknown for its base model Qwen3: training data not disclosed in enough detail
Notes
No benchmark is named as training data and no decontamination statement against multimodal benchmarks was found in the report. Evaluations report ChartQA/DocVQA/AI2D test, MMMU, MathVista testmini, CharXiv. Accepts image input: “The Qwen3-VL framework integrates a vision encoder and a language model decoder to process multimodal inputs, including text, images, and video.”
Sources
- Qwen3-VL Technical Report (arXiv 2511.21631v2) · paper · retrieved 2026-10-15
- Qwen3-VL-8B-Instruct model card · model card · retrieved 2026-10-15