Contamination Map Methodology

Text / LLM / Model

K2-V2

Organisation
LLM360 / MBZUAI Institute of Foundation Models (IFM)
Family
LLM360 K2 · version 70B
Released
2025-12-05 (arXiv v1 submission date of 2512.06201 (https://arxiv.org/abs/2512.06201), also the release date stated on the TxT360-Midas/TxT360-3efforts cards; HF createdAt IFM/K2-V2 2025-12-01)
Checkpoints
IFM/K2-V2, IFM/K2-V2-Instruct
Weights
Open · apache-2.0
Links
PaperModel cardCode

Training data disclosure: Full list

Level determined for 6% of benchmark cells.

“All mixtures, filtering rules, and data sources are fully released for reproducibility.”
huggingface.co · model card · retrieved 2026-10-15
“An overview of the complete pre-training data is presented in Table 1.”
arxiv.org · paper · Section 3.3 · retrieved 2026-10-15
“By fully releasing the training data, K2 allows the community to gain a deeper understanding of these dependencies”
arxiv.org · paper · Section 1 · retrieved 2026-10-15

Training corpora

Other declared training data

Contamination analyses

Analyses that found benchmark test data in this model's training data, or in a corpus it uses.

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

Test data seen AIME 2024 · a contamination analysis found test examples in the pre-training data (Sentence-level memorisation rate 23.32% for K2 base (beam-search continuation exactly reproduces the next sentence of the question); lower than Qwen2.5-72B (41.96%) and Olmo 3 32B (37.82%); authors attribute AIME performance to reasoning rather than memorisation.)

Evidence (2)
“we conduct memorization analysis on AIME 2024 and AIME 2025 questions.”
arxiv.org · paper · Section 4.2, Memorization analysis · retrieved 2026-10-15
“Table 4: Sentence-level memorization rates (%).”
arxiv.org · paper · Table 4 · retrieved 2026-10-15

Unknown AIME 2025 · training data not disclosed in enough detail

Unknown MATH-500 · training data not disclosed in enough detail

Notes

70B dense; ~12T-token pre-training, four mid-training stages (TxT360-Midas, recorded as pretraining) and SFT only (no RL/preference optimisation per report). Table 1 of the report gives per-source information cut-off dates (e.g. CommonCrawl July 2024, MegaMath Nov 2024) but no model-level training cutoff is stated. Jais Arabic data is named but not shown to be published separately. Decontamination statements not shown as cells (they name no benchmark tracked here, or none at all): 12 downstream benchmarks (MegaMath decontamination; benchmarks not named in K2-V2 report); evaluation benchmarks (unspecified) - SFT data; test subsets of mid-training math problem sources. Other reported contamination, not shown as cells: aime-2025: Sentence-level memorisation rate 14.43% for K2 base; the authors say most 2025 memorisation comes from the last, formulaic sentence of each problem.

Sources