Contamination MapText / LLM / Model
Gemma 2
- Organisation
- Google DeepMind
- Family
- Gemma · version 2
- Released
- 2024-06-27 (Google blog 'Gemma 2 is now available to researchers and developers' dated Jun 27, 2024 (9B and 27B); the 2B followed later (HF repo created 2024-07-16))
- Checkpoints
- google/gemma-2-2b, google/gemma-2-9b, google/gemma-2-27b, google/gemma-2-2b-it, google/gemma-2-9b-it, google/gemma-2-27b-it
- Weights
- Open · Gemma Terms of Use
- 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
- ?Unknown15
“These tokens come from a variety of data sources, including web documents, code, and science articles.”
“These models were trained on a dataset of text data that includes a wide variety of sources.”
Other declared training data
-
web documents (broad web data: may contain benchmark test sets) · split: unknown
27B: 13T tokens, 9B: 8T, 2B: 2T, primarily English.
Evidence
“These tokens come from a variety of data sources, including web documents, code, and science articles.”
arxiv.org · paper · Section 3.1 Training Data · retrieved 2026-10-15 “A diverse collection of web text ensures the model is exposed to a broad range of linguistic styles, topics, and vocabulary.”
ai.google.dev · model card · Model Data / Training Dataset · retrieved 2026-10-15 -
code (not yet matched to a dataset) · split: unknown
Source not described.
Evidence
“These tokens come from a variety of data sources, including web documents, code, and science articles.”
arxiv.org · paper · Section 3.1 Training Data · retrieved 2026-10-15 -
science articles (not yet matched to a dataset) · split: unknown
Evidence
“These tokens come from a variety of data sources, including web documents, code, and science articles.”
arxiv.org · paper · Section 3.1 Training Data · retrieved 2026-10-15 -
mathematics (not yet matched to a dataset) · split: unknown
Evidence
“Training on mathematical text helps the model learn logical reasoning, symbolic representation, and to address mathematical queries.”
ai.google.dev · model card · Model Data / Training Dataset · retrieved 2026-10-15 -
internal and external public post-training data (broad web data: may contain benchmark test sets) · split: unknown · fine-tuning
Not named beyond LMSYS-chat-1M prompts.
Evidence
“We extended the post-training data from Gemma 1.1 with a mixture of internal and external public data.”
arxiv.org · paper · Section 4 Post-Training · retrieved 2026-10-15 -
LMSYS-chat-1M (prompts only) (not yet matched to a dataset) · split: unknown · fine-tuning
User prompts collected from Chatbot Arena; may contain benchmark questions pasted by users (not analysed by the authors).
Evidence
“In particular, we use the prompts, but not the answers from LMSYS-chat-1M (Zheng et al., 2023).”
arxiv.org · paper · Section 4 Post-Training · retrieved 2026-10-15 -
synthetic responses generated by a larger teacher model (not yet matched to a dataset) · split: unknown · fine-tuning
Evidence
“We run behavioral cloning on synthetic and real prompts, and responses predominantly synthetically generated by the teacher, that is a larger model.”
arxiv.org · paper · Section 4 Post-Training, Supervised fine-tuning · retrieved 2026-10-15
Benchmark cells
Classic LLM benchmarks
Unknown MMLU (Measuring Massive Multitask Language Understanding) · training data not disclosed in enough detail
Unknown GSM8K (Grade School Math 8K) · training data not disclosed in enough detail
Unknown MATH (Mathematics Aptitude Test of Heuristics) · training data not disclosed in enough detail
Unknown HumanEval · training data not disclosed in enough detail
Unknown ARC-Challenge (AI2 Reasoning Challenge, Challenge Set) · training data not disclosed in enough detail
Unknown HellaSwag · training data not disclosed in enough detail
Unknown TruthfulQA · training data not disclosed in enough detail
Unknown WinoGrande · training data not disclosed in enough detail
Recent LLM benchmarks
Unknown GPQA Diamond · training data not disclosed in enough detail
Unknown MMLU-Pro · training data not disclosed in enough detail
Unknown LiveCodeBench (code generation, release_v6) · training data not disclosed in enough detail
Unknown SWE-bench Verified · training data not disclosed in enough detail
Unknown AIME 2024 · training data not disclosed in enough detail
Unknown AIME 2025 · training data not disclosed in enough detail
Unknown MATH-500 · training data not disclosed in enough detail
Notes
No knowledge cutoff stated in the paper or model card. The paper's memorization analysis concerns training-data recitation, not benchmarks. Decontamination statements not shown as cells (they name no benchmark tracked here, or none at all): evaluation sets (not named).
Sources
- Gemma 2: Improving Open Language Models at a Practical Size (arXiv v3) · paper · retrieved 2026-10-15
- Gemma 2 model card (ai.google.dev) · model card · retrieved 2026-10-15
- Gemma 2 launch blog (Google) · website · retrieved 2026-10-15