Contamination Map Methodology

Text / LLM / Model

InternVL3

Organisation
Shanghai AI Laboratory (OpenGVLab)
Family
InternVL · version 3
Released
2025-04-10 (HF repo creation of OpenGVLab/InternVL3-*B (2025-04-10, via huggingface.co API); arXiv v1 2025-04-14)
Checkpoints
OpenGVLab/InternVL3-1B, OpenGVLab/InternVL3-2B, OpenGVLab/InternVL3-8B, OpenGVLab/InternVL3-9B, OpenGVLab/InternVL3-14B, OpenGVLab/InternVL3-38B, OpenGVLab/InternVL3-78B, OpenGVLab/InternVL3-8B-Instruct, OpenGVLab/InternVL3-8B-Pretrained
Weights
Open · Apache 2.0 (model card also links the Qwen license for the Qwen2.5-based components)
Links
PaperModel cardCode

Training data disclosure: Partial list

Level determined for 31% of benchmark cells.

“The open-source data are being organized, and a comprehensive list will be included in a future revision of this report.”
arxiv.org · paper · Footnote 1, page 3 · retrieved 2026-10-15
“For SFT data, we construct the training corpora based on those used in InternVL2.5 [18] while introducing additional tool usage, 3D scene understanding, GUI operations, scientific diagrams, creative writing, and multimodal reasoning samples.”
arxiv.org · paper · Section 2.3 Data · retrieved 2026-10-15
“We originally intended to open-source the training dataset used in InternVL. However, due to the company policy, we have withdrawn the dataset from public access.”
huggingface.co · dataset card · Note on Dataset Availability · retrieved 2026-10-15

Built on

Everything these base models may have seen is inherited by InternVL3.

Other declared training data

Benchmark cells

Classic LLM benchmarks

Unknown MMLU (Measuring Massive Multitask Language Understanding) · unknown for its base model Qwen2.5: training data not disclosed in enough detail

Unknown GSM8K (Grade School Math 8K) · unknown for its base model Qwen2.5: training data not disclosed in enough detail

Related data seen MATH (Mathematics Aptitude Test of Heuristics) · via base model Qwen2.5: a contamination analysis found test examples of the related dataset MATH-500 (Exact completion of 21.2% of MATH-500 problems from a 60% prefix (Table 2).) [third-party study]

Evidence (4)
“For the language model, we leverage pre-trained large language models (LLMs), specifically the Qwen2.5 series and InternLM3-8B.”
arxiv.org · paper · Section 2.1 · retrieved 2026-10-15
“revealing strong signs of data contamination in the Qwen2.5 series models when evaluated on commonly used benchmarks, such as MATH-500, AMC, and AIME2024.”
arxiv.org third-party study· study · Section 4, Table 2 · retrieved 2026-10-15
“We therefore evaluate our models only on the remaining 500 held-out problems. We selected these 500 test problems uniformly at random, and we believe they are representative of the test set as a whole.”
raw.githubusercontent.com · readme · MATH Splits · retrieved 2026-10-15
“This dataset contains a subset of 500 problems from the MATH benchmark that OpenAI created in their _Let's Verify Step by Step_ paper.”
huggingface.co · dataset card · retrieved 2026-10-15

Unknown HumanEval · unknown for its base model Qwen2.5: training data not disclosed in enough detail

Unknown ARC-Challenge (AI2 Reasoning Challenge, Challenge Set) · unknown for its base model Qwen2.5: training data not disclosed in enough detail

Unknown HellaSwag · unknown for its base model Qwen2.5: training data not disclosed in enough detail

Unknown TruthfulQA · unknown for its base model Qwen2.5: training data not disclosed in enough detail

Unknown WinoGrande · unknown for its base model Qwen2.5: training data not disclosed in enough detail

Multimodal LLM benchmarks

Unknown MMMU (Massive Multi-discipline Multimodal Understanding and Reasoning) · unknown for its base model Qwen2.5: training data not disclosed in enough detail

Unknown MMMU-Pro · unknown for its base model Qwen2.5: training data not disclosed in enough detail

Related data seen MathVista · related dataset ChartQA was used

Evidence (5)
“Specifically, we leverage the pre-training corpus from InternVL2.5, which covers a diverse range of domains such as image captioning, general question answering, mathematics, charts, optical character recognition (OCR), knowledge grounding, document understanding, multi-turn dialogue, and medical data.”
arxiv.org · paper · Section 2.2 Data · retrieved 2026-10-15
“ChartQA [181], PlotQA [187], FigureQA [105], LRV-Instruction [148], ArxivQA [132], MMC-Inst [149]”
arxiv.org · paper · Table 4 (InternVL 2.5 pre-training data mixture), Chart row · retrieved 2026-10-15
“It consists of 6,141 examples, derived from 28 existing multimodal datasets involving mathematics and 3 newly created datasets”
arxiv.org · paper · Abstract · retrieved 2026-10-15
“ChartQA (2022)”
arxiv.org · paper · Table 5 (Appendix C) · retrieved 2026-10-15
“"dataset": "ChartQA",”
huggingface.co · dataset card · source.json · retrieved 2026-10-15

Seen, split unclear ChartQA · the dataset was used; the source does not say which part

Evidence (2)
“Specifically, we leverage the pre-training corpus from InternVL2.5, which covers a diverse range of domains such as image captioning, general question answering, mathematics, charts, optical character recognition (OCR), knowledge grounding, document understanding, multi-turn dialogue, and medical data.”
arxiv.org · paper · Section 2.2 Data · retrieved 2026-10-15
“ChartQA [181], PlotQA [187], FigureQA [105], LRV-Instruction [148], ArxivQA [132], MMC-Inst [149]”
arxiv.org · paper · Table 4 (InternVL 2.5 pre-training data mixture), Chart row · retrieved 2026-10-15

Seen, split unclear DocVQA · the dataset was used; the source does not say which part

Evidence (2)
“Specifically, we leverage the pre-training corpus from InternVL2.5, which covers a diverse range of domains such as image captioning, general question answering, mathematics, charts, optical character recognition (OCR), knowledge grounding, document understanding, multi-turn dialogue, and medical data.”
arxiv.org · paper · Section 2.2 Data · retrieved 2026-10-15
“DocReason25K [93], DocVQA [184], Docmatix [121], Synthetic Arxiv QA”
arxiv.org · paper · Table 4 (InternVL 2.5 pre-training data mixture), Document row · retrieved 2026-10-15

Seen, split unclear AI2D (AI2 Diagrams) · the dataset was used; the source does not say which part

Evidence (2)
“Specifically, we leverage the pre-training corpus from InternVL2.5, which covers a diverse range of domains such as image captioning, general question answering, mathematics, charts, optical character recognition (OCR), knowledge grounding, document understanding, multi-turn dialogue, and medical data.”
arxiv.org · paper · Section 2.2 Data · retrieved 2026-10-15
“KonIQ-10K [91], IconQA [167], VisualMRC [225], ChemVLM Data [129], ScienceQA [165], AI2D [109]”
arxiv.org · paper · Table 4 (InternVL 2.5 pre-training data mixture), Knowledge row · retrieved 2026-10-15

Unknown CharXiv · unknown for its base model Qwen2.5: training data not disclosed in enough detail

Recent LLM benchmarks

Unknown GPQA Diamond · unknown for its base model Qwen2.5: training data not disclosed in enough detail

Unknown MMLU-Pro · unknown for its base model Qwen2.5: training data not disclosed in enough detail

Unknown LiveCodeBench (code generation, release_v6) · unknown for its base model Qwen2.5: training data not disclosed in enough detail

Unknown SWE-bench Verified · unknown for its base model Qwen2.5: training data not disclosed in enough detail

Test data seen AIME 2024 · via base model Qwen2.5: a contamination analysis found test examples in the pre-training data (Exact completion of 13.33% of AIME 2024 problems from a 60% prefix (Table 2).) [third-party study]

Evidence (2)
“For the language model, we leverage pre-trained large language models (LLMs), specifically the Qwen2.5 series and InternLM3-8B.”
arxiv.org · paper · Section 2.1 · retrieved 2026-10-15
“revealing strong signs of data contamination in the Qwen2.5 series models when evaluated on commonly used benchmarks, such as MATH-500, AMC, and AIME2024.”
arxiv.org third-party study· study · Section 4, Table 2 · retrieved 2026-10-15

Unknown AIME 2025 · unknown for its base model Qwen2.5: training data not disclosed in enough detail

Test data seen MATH-500 · via base model Qwen2.5: a contamination analysis found test examples in the pre-training data (Exact completion of 21.2% of MATH-500 problems from a 60% prefix (Table 2).) [third-party study]

Evidence (2)
“For the language model, we leverage pre-trained large language models (LLMs), specifically the Qwen2.5 series and InternLM3-8B.”
arxiv.org · paper · Section 2.1 · retrieved 2026-10-15
“revealing strong signs of data contamination in the Qwen2.5 series models when evaluated on commonly used benchmarks, such as MATH-500, AMC, and AIME2024.”
arxiv.org third-party study· study · Section 4, Table 2 · retrieved 2026-10-15

Notes

The promised full data list was not published: the InternVL3 v3 report still says it 'will be included in a future revision', and the OpenGVLab/InternVL-Data release was withdrawn. Benchmark-named data is known only through the InternVL2.5 tables that InternVL3 says it builds on. No decontamination statement against multimodal benchmarks found. Accepts image input: “The key difference, starting from InternVL 2.0, is that we additionally introduced support for multi-image and video data.”

Sources