Contamination MapText / LLM / Model
DeepSeek-V3
- Organisation
- DeepSeek-AI
- Family
- DeepSeek · version V3
- Released
- 2024-12-26 (DeepSeek API Docs news entry 'DeepSeek-V3 2024/12/26', https://api-docs.deepseek.com/news/news1226)
- Training cutoff
- 2024-07-01
Evidence
“DeepSeek-V3 base has a knowledge cutoff date of July 2024, predating evaluation benchmarks like CNMO 2024,”
arxiv.org · paper · Appendix D.1 (Decontamination) · retrieved 2026-10-15 - Checkpoints
- deepseek-ai/DeepSeek-V3-Base, deepseek-ai/DeepSeek-V3
- Weights
- Open · MIT (code); DeepSeek Model License (weights, commercial use allowed)
- Links
- PaperModel cardCode
Training data disclosure: Vague
Level determined for 6% of benchmark cells.
- TestTest data seen0
- SeenSeen, split unclear0
- TrainTraining data seen0
- Related data seen0
- NoNot seen1
- ?Unknown14
“Finally, the training corpus for DeepSeek-V3 consists of 14.8T high-quality and diverse tokens in our tokenizer.”
“Compared with DeepSeek-V2, we optimize the pre-training corpus by enhancing the ratio of mathematical and programming samples, while expanding multilingual coverage beyond”
Other declared training data
-
plain web pages and e-books (DeepSeek-V3-Base pre-training data, 14.8T tokens) (broad web data: may contain benchmark test sets) · split: unknown
Described in the DeepSeek-R1 paper (same developer). The V3 report adds that math and programming samples were up-weighted.
Evidence
“For the training data of DeepSeek-V3-Base, we exclusively use plain web pages and e-books, without incorporating any synthetic data.”
arxiv.org · paper · Appendix A.1 · retrieved 2026-10-15 “all data used in this phase were naturally occurring and collected through web crawling.”
arxiv.org · paper · Appendix A.1 · retrieved 2026-10-15 -
instruction-tuning datasets (1.5M instances) (not yet matched to a dataset) · split: unknown · fine-tuning
Evidence
“We curate our instruction-tuning datasets to include 1.5M instances spanning multiple domains, with each domain employing distinct data creation methods tailored to its specific requirements.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15 -
reasoning SFT data generated by an internal DeepSeek-R1 model (math, code competition, logic puzzles) (not yet matched to a dataset) · split: unknown · fine-tuning
Evidence
“For reasoning-related datasets, including those focused on mathematics, code competition problems, and logic puzzles, we generate the data by leveraging an internal DeepSeek-R1 model.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15 -
non-reasoning SFT data generated by DeepSeek-V2.5 and human-verified (not yet matched to a dataset) · split: unknown · fine-tuning
Evidence
“For non-reasoning data, such as creative writing, role-play, and simple question answering, we utilize DeepSeek-V2.5 to generate responses and enlist human annotators to verify the accuracy and correctness of the data.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
Declared exclusions
-
Dataset MMLU (Measuring Massive Multitask Language Understanding) · test portion excluded from pre-training data
Evidence
“we filtered out any text segments (including web pages and GitHub files) that contained matching 10-gram sequences from evaluation questions or reference solutions.”
arxiv.org · paper · Appendix D.1 (Decontamination) · retrieved 2026-10-15 “Benchmarks We evaluate models on MMLU (Hendrycks et al., 2021), MMLU-Redux (Gema et al., 2025), MMLU-Pro (Wang et al., 2024),”
arxiv.org · paper · Appendix D.1 (Benchmarks) · retrieved 2026-10-15 -
Dataset MMLU-Pro · test portion excluded from pre-training data
Evidence
“we filtered out any text segments (including web pages and GitHub files) that contained matching 10-gram sequences from evaluation questions or reference solutions.”
arxiv.org · paper · Appendix D.1 (Decontamination) · retrieved 2026-10-15 “Benchmarks We evaluate models on MMLU (Hendrycks et al., 2021), MMLU-Redux (Gema et al., 2025), MMLU-Pro (Wang et al., 2024),”
arxiv.org · paper · Appendix D.1 (Benchmarks) · retrieved 2026-10-15 -
Dataset GPQA Diamond · test portion excluded from pre-training data
Evidence
“we filtered out any text segments (including web pages and GitHub files) that contained matching 10-gram sequences from evaluation questions or reference solutions.”
arxiv.org · paper · Appendix D.1 (Decontamination) · retrieved 2026-10-15 “GPQA Diamond (Rein et al., 2023), SimpleQA (OpenAI, 2024a), C-SimpleQA (He et al., 2024), SWE-Bench Verified (OpenAI, 2024b),”
arxiv.org · paper · Appendix D.1 (Benchmarks) · retrieved 2026-10-15 -
Dataset SWE-bench Verified · test portion excluded from pre-training data
Evidence
“we filtered out any text segments (including web pages and GitHub files) that contained matching 10-gram sequences from evaluation questions or reference solutions.”
arxiv.org · paper · Appendix D.1 (Decontamination) · retrieved 2026-10-15 “GPQA Diamond (Rein et al., 2023), SimpleQA (OpenAI, 2024a), C-SimpleQA (He et al., 2024), SWE-Bench Verified (OpenAI, 2024b),”
arxiv.org · paper · Appendix D.1 (Benchmarks) · retrieved 2026-10-15 -
Dataset LiveCodeBench (code generation, release_v6) · test portion excluded from pre-training data
Evidence
“we filtered out any text segments (including web pages and GitHub files) that contained matching 10-gram sequences from evaluation questions or reference solutions.”
arxiv.org · paper · Appendix D.1 (Decontamination) · retrieved 2026-10-15 “LiveCodeBench (Jain et al., 2024)”
arxiv.org · paper · Appendix D.1 (Benchmarks) · retrieved 2026-10-15 -
Dataset AIME 2024 · test portion excluded from pre-training data
Evidence
“we filtered out any text segments (including web pages and GitHub files) that contained matching 10-gram sequences from evaluation questions or reference solutions.”
arxiv.org · paper · Appendix D.1 (Decontamination) · retrieved 2026-10-15 “American Invitational Mathematics Examination 2024 (AIME 2024) (MAA, 2024).”
arxiv.org · paper · Appendix D.1 (Benchmarks) · retrieved 2026-10-15 -
Dataset MATH-500 · test portion excluded from pre-training data
Evidence
“we filtered out any text segments (including web pages and GitHub files) that contained matching 10-gram sequences from evaluation questions or reference solutions.”
arxiv.org · paper · Appendix D.1 (Decontamination) · retrieved 2026-10-15 “Finally, AIME, MATH-500, and CNMO 2024 comprise mathematics problems that test the model”
arxiv.org · paper · Appendix D.1 (Benchmarks) · 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
Not seen AIME 2025 · its test items were published from 2025-02-06, after the model's stated training-data cutoff (2024-07-01)
Evidence (3)
“DeepSeek-V3 base has a knowledge cutoff date of July 2024, predating evaluation benchmarks like CNMO 2024,”
“2025 AIME I problems and solutions. The test was held on Thursday, February 6, 2025.”
“This year the AIME I will take place on February 6, 2025, while the alternate AIME II will take place on February 12, 2025.”
Unknown MATH-500 · training data not disclosed in enough detail
Notes
Pre-training: 14.8T tokens. Cutoff, pre-training source types and decontamination come from the DeepSeek-R1 paper (v2, Jan 2026), which describes DeepSeek-V3-Base. Other reported contamination, not shown as cells: evaluation benchmarks (mathematics domain): 10-gram decontamination identified and removed about six million potential pre-training texts in the mathematics domain alone (before training)
Sources
- DeepSeek-V3 Technical Report · paper · retrieved 2026-10-15
- DeepSeek-V3 model card · model card · retrieved 2026-10-15
- DeepSeek-R1 (v2, with appendices on DeepSeek-V3-Base data) · paper · retrieved 2026-10-15
- DeepSeek API Docs: Introducing DeepSeek-V3 · website · retrieved 2026-10-15