Contamination MapText / LLM / Model
Olmo 3.1
- Organisation
- Allen Institute for AI (Ai2)
- Family
- OLMo · version 3.1
- Released
- 2025-12-12 ('Update 12/12: Announcing Olmo 3.1' on the Olmo 3 blog (https://allenai.org/blog/olmo3); HF repos created 2025-12-10 (HF API))
- Training cutoff
- 2024-12-01
Evidence
“Date cutoff:** Dec. 2024.”
huggingface.co · model card · Model Description · retrieved 2026-10-15 - Checkpoints
- allenai/Olmo-3.1-32B-Think, allenai/Olmo-3.1-32B-Instruct-SFT, allenai/Olmo-3.1-32B-Instruct-DPO, allenai/Olmo-3.1-32B-Instruct, allenai/Olmo-3.1-7B-RL-Zero-Math, allenai/Olmo-3.1-7B-RL-Zero-Code
- Weights
- Open · apache-2.0
- Links
- PaperModel cardCode
Training data disclosure: Full list
Level determined for 6% of benchmark cells.
- TestTest data seen0
- SeenSeen, split unclear0
- TrainTraining data seen0
- Related data seen0
- NoNot seen1
- ?Unknown14
“These models are pre-trained on the Dolma 3 dataset and post-trained on the Dolci datasets.”
“we resumed our RL training run for Olmo 3 32B Think, training for an additional 21 days on 224 GPUs with extra epochs over our Dolci-Think-RL dataset.”
Training corpora
-
Dolma 3 Mix
· 6 sources (complete list) · pre-training
· homepage
Evidence that Olmo 3.1 uses it
“Dataset: [dolma3_mix-5.5T-1125](https://huggingface.co/datasets/allenai/dolma3_mix-5.5T-1125)”
huggingface.co · model card · Stage 1 · retrieved 2026-10-15 Contents of this corpus
-
common_crawl (broad web data: may contain benchmark test sets) · split: unknown
Evidence
“common_crawl | web pages | 4.51T”
huggingface.co · dataset card · Source Sizes · retrieved 2026-10-15 “Text extraction We start with 104 dumps from the CommonCrawl corpus, with a cutoff date of December”
arxiv.org · paper · Section 3.4 · retrieved 2026-10-15 -
olmocr_science_pdfs (broad web data: may contain benchmark test sets) · split: unknown
Crawled PDFs (238M unique PDFs, cutoff Dec 2024) converted with olmOCR.
Evidence
“olmocr_science_pdfs | academic papers | 805B”
huggingface.co · dataset card · Source Sizes · retrieved 2026-10-15 -
stack_edu (broad web data: may contain benchmark test sets) · split: unknown
Evidence
“stack_edu | code | 409B”
huggingface.co · dataset card · Source Sizes · retrieved 2026-10-15 “For code data, we use Stack-Edu (Allal et al., 2025), an improved curation of GitHub repositories from”
arxiv.org · paper · Section 3.4 · retrieved 2026-10-15 -
finemath-3plus (broad web data: may contain benchmark test sets) · split: unknown
FineMath: Common Crawl math pages.
Evidence
“finemath-3plus | mathematics | 151B”
huggingface.co · dataset card · Source Sizes · retrieved 2026-10-15 “This data has a cutoff date of September 2024.”
arxiv.org · paper · Section 3.4 · retrieved 2026-10-15 -
rpj-proofpile-arxiv (not yet matched to a dataset) · split: unknown
Evidence
“rpj-proofpile-arxiv | research papers | 50.9B”
huggingface.co · dataset card · Source Sizes · retrieved 2026-10-15 -
dolma1_7-wiki-en (not yet matched to a dataset) · split: unknown
Evidence
“dolma1_7-wiki-en | encyclopedic | 2.51B”
huggingface.co · dataset card · Source Sizes · retrieved 2026-10-15
-
common_crawl (broad web data: may contain benchmark test sets) · split: unknown
-
Dolma 3 Dolmino Mix
· 24 sources (complete list) · pre-training
· homepage
Evidence that Olmo 3.1 uses it
“Dataset: [dolma3-dolmino-mix-1125](https://huggingface.co/datasets/allenai/dolma3_dolmino_mix-100B-1125)”
huggingface.co · model card · Stage 2 · retrieved 2026-10-15 “For decontamination, we search for and remove matches of any split of any benchmark”
arxiv.org · paper · Section 3.5.3 · retrieved 2026-10-15 “We focus our decontamination efforts on the midtraining stage (and the long-context extension, which drew from the same data pools)”
arxiv.org · paper · Section 3.5.3 · retrieved 2026-10-15 Contents of this corpus
-
TinyMATH Mind (not yet matched to a dataset) · split: unknown
Synthetic problems generated from the MATH training set.
Evidence
“TinyMATH Mind | Math (synth) | 898M (0.9%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 “TinyMATH For each of the 7500 examples in the MATH training set, we generate 100 new, similar”
arxiv.org · paper · Section 3.5 · retrieved 2026-10-15 -
TinyMATH PoT (not yet matched to a dataset) · split: unknown
Synthetic, derived from the MATH training set.
Evidence
“TinyMATH PoT | Math (synth) | 241M (0.24%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
CraneMath (not yet matched to a dataset) · split: unknown
Evidence
“CraneMath | Math (synth) | 5.62B (5.63%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
MegaMatt (not yet matched to a dataset) · split: unknown
Evidence
“MegaMatt | Math (synth) | 1.73B (1.73%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
Dolmino Math (not yet matched to a dataset) · split: unknown
OLMo 2 Dolmino math subset (which contained GSM8K train and GSM8K-derived data), re-filtered by Olmo 3 decontamination, which removes matches of any benchmark split.
Evidence
“Dolmino Math | Math (synth) | 10.7B (10.7%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 “The version we use differs from the original only in additional filtering for decontamination.”
arxiv.org · paper · Section 3.5 · retrieved 2026-10-15 -
StackEdu (FIM) (broad web data: may contain benchmark test sets) · split: unknown
Evidence
“StackEdu (FIM) | Code | 10.0B (10.0%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
CraneCode (not yet matched to a dataset) · split: unknown
Evidence
“CraneCode | Python (synth) | 10.0B (10.0%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
Reddit To Flashcards (not yet matched to a dataset) · split: unknown
Evidence
“Reddit To Flashcards | QA (synth) | 5.90B (5.9%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
Wiki To RCQA (not yet matched to a dataset) · split: unknown
Evidence
“Wiki To RCQA | QA (synth) | 3.0B (3.0%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
Nemotron Synth QA (not yet matched to a dataset) · split: unknown
Evidence
“Nemotron Synth QA | QA (synth) | 5.0B (5.0%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
Math Meta-Reasoning (not yet matched to a dataset) · split: unknown
Evidence
“Math Meta-Reasoning | Thinking (synth) | 381M (0.38%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
Code Meta-Reasoning (not yet matched to a dataset) · split: unknown
Evidence
“Code Meta-Reasoning | Thinking (synth) | 459M (0.46%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
Program-Verifiable (not yet matched to a dataset) · split: unknown
Evidence
“Program-Verifiable | Thinking (synth) | 159M (0.16%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
OMR Rewrite FullThoughts (not yet matched to a dataset) · split: unknown
Evidence
“OMR Rewrite FullThoughts | Thinking (synth) | 850M (0.85%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
QWQ Reasoning Traces (not yet matched to a dataset) · split: unknown
Evidence
“QWQ Reasoning Traces | Thinking (synth) | 1.87B (1.87%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
General Reasoning Mix (not yet matched to a dataset) · split: unknown
Evidence
“General Reasoning Mix | Thinking (synth) | 1.87B (1.87%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
Gemini Reasoning Traces (not yet matched to a dataset) · split: unknown
Evidence
“Gemini Reasoning Traces | Thinking (synth) | 246M (0.25%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
Llama Nemotron Reasoning Traces (not yet matched to a dataset) · split: unknown
Evidence
“Llama Nemotron Reasoning Traces | Thinking (synth) | 1.25B (1.25%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
OpenThoughts2 Reasoning Traces (not yet matched to a dataset) · split: unknown
Evidence
“OpenThoughts2 Reasoning Traces | Thinking (synth) | 1.25B (1.25%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
Tulu 3 SFT (not yet matched to a dataset) · split: unknown
Evidence
“Tulu 3 SFT | Instruction (synth) | 1.1B (1.1%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
Dolmino 1 Flan (not yet matched to a dataset) · split: unknown
Evidence
“Dolmino 1 Flan | Instruction (synth) | 5.0B (5.0%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
OLMOCR Science PDFs (High Q.) (broad web data: may contain benchmark test sets) · split: unknown
Evidence
“OLMOCR Science PDFs (High Q.) | PDFs | 4.99B (5.0%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
STEM-Heavy Crawl (broad web data: may contain benchmark test sets) · split: unknown
Evidence
“STEM-Heavy Crawl | Web pages | 4.99B (5.0%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
Common Crawl (High Q.) (broad web data: may contain benchmark test sets) · split: unknown
Evidence
“Common Crawl (High Q.) | Web pages | 22.4B (22.5%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15
-
TinyMATH Mind (not yet matched to a dataset) · split: unknown
-
Dolma 3 Longmino Mix
· 2 sources (complete list) · pre-training
· homepage
Evidence that Olmo 3.1 uses it
“Dataset: [dolma3-longmino-mix-1125](https://huggingface.co/datasets/allenai/dolma3_longmino_mix-100B-1125)”
huggingface.co · model card · Stage 3 · retrieved 2026-10-15 Contents of this corpus
-
LC-s2pdf (long PDFs, incl. synthetic REX/CWE variants) (broad web data: may contain benchmark test sets) · split: unknown
Evidence
“LC-s2pdf 32k-64k | PDFs | 4.81B (9.63%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15 -
Midtraining Data (not yet matched to a dataset) · split: unknown
Drawn from the Dolma 3 Dolmino pool.
Evidence
“Midtraining Data | Mix | 33.0B (66.1%)”
huggingface.co · dataset card · Dataset Sources · retrieved 2026-10-15
-
LC-s2pdf (long PDFs, incl. synthetic REX/CWE variants) (broad web data: may contain benchmark test sets) · split: unknown
Other declared training data
-
Dolci-Think-SFT / Dolci-Think-DPO (not yet matched to a dataset) · split: unknown · fine-tuning
Olmo 3.1 Think 32B continues from Olmo 3 32B Think (same SFT/DPO data).
Evidence
“supervised fine-tuning on the Dolci-Think-SFT-7B dataset. This dataset consits of math, code, chat, and general knowledge queries.”
huggingface.co · model card · Post-training · retrieved 2026-10-15 -
Dolci-Think-RL (not yet matched to a dataset) · split: unknown · fine-tuning
Evidence
“with extra epochs over our Dolci-Think-RL dataset”
allenai.org · website · Update 12/12 · retrieved 2026-10-15 -
Dolci-Instruct-SFT / DPO / RL (not yet matched to a dataset) · split: unknown · fine-tuning
Evidence
“Olmo 3.1 Instruct 32B, which applies the recipe behind Olmo 3 Instruct 7B to a much larger model”
allenai.org · website · Update 12/12 · retrieved 2026-10-15 “allenai/Dolci-Instruct-RL”
huggingface.co · model card · Model card metadata (datasets) · retrieved 2026-10-15
Declared exclusions
-
Dataset MATH-500 · test portion excluded from fine-tuning data
Evidence
“to filter out the portions of all post-training data (all three stages) that matched the evaluation sets.”
arxiv.org · paper · Section 4.2.1, Step 4 · retrieved 2026-10-15 “We used n-gram matching with 8-grams and an overlap threshold of 0.5”
arxiv.org · paper · Section 4.2.1, Step 4 · retrieved 2026-10-15 “MATH 500 (2022; 2023)”
arxiv.org · paper · Table 16 (Olmo 3 chat evaluation suite) · retrieved 2026-10-15 -
Dataset AIME 2024 · test portion excluded from fine-tuning data
Evidence
“to filter out the portions of all post-training data (all three stages) that matched the evaluation sets.”
arxiv.org · paper · Section 4.2.1, Step 4 · retrieved 2026-10-15 “We used n-gram matching with 8-grams and an overlap threshold of 0.5”
arxiv.org · paper · Section 4.2.1, Step 4 · retrieved 2026-10-15 “AIME 2024*”
arxiv.org · paper · Table 16 (Olmo 3 chat evaluation suite) · retrieved 2026-10-15 -
Dataset AIME 2025 · test portion excluded from fine-tuning data
Evidence
“to filter out the portions of all post-training data (all three stages) that matched the evaluation sets.”
arxiv.org · paper · Section 4.2.1, Step 4 · retrieved 2026-10-15 “We used n-gram matching with 8-grams and an overlap threshold of 0.5”
arxiv.org · paper · Section 4.2.1, Step 4 · retrieved 2026-10-15 “AIME 2025*”
arxiv.org · paper · Table 16 (Olmo 3 chat evaluation suite) · retrieved 2026-10-15 -
Dataset LiveCodeBench (code generation, release_v6) · test portion excluded from fine-tuning data
Evidence
“to filter out the portions of all post-training data (all three stages) that matched the evaluation sets.”
arxiv.org · paper · Section 4.2.1, Step 4 · retrieved 2026-10-15 “We used n-gram matching with 8-grams and an overlap threshold of 0.5”
arxiv.org · paper · Section 4.2.1, Step 4 · retrieved 2026-10-15 “LiveCodeBench v3* (2024)”
arxiv.org · paper · Table 16 (Olmo 3 chat evaluation suite) · retrieved 2026-10-15 -
Dataset GPQA (Graduate-Level Google-Proof Q&A) · test portion excluded from fine-tuning data
Evidence
“to filter out the portions of all post-training data (all three stages) that matched the evaluation sets.”
arxiv.org · paper · Section 4.2.1, Step 4 · retrieved 2026-10-15 “We used n-gram matching with 8-grams and an overlap threshold of 0.5”
arxiv.org · paper · Section 4.2.1, Step 4 · retrieved 2026-10-15 “GPQA* (2024)”
arxiv.org · paper · Table 16 (Olmo 3 chat evaluation suite) · retrieved 2026-10-15 -
Dataset MMLU (Measuring Massive Multitask Language Understanding) · test portion excluded from fine-tuning data
Evidence
“to filter out the portions of all post-training data (all three stages) that matched the evaluation sets.”
arxiv.org · paper · Section 4.2.1, Step 4 · retrieved 2026-10-15 “We used n-gram matching with 8-grams and an overlap threshold of 0.5”
arxiv.org · paper · Section 4.2.1, Step 4 · retrieved 2026-10-15 “MMLU (2021b)”
arxiv.org · paper · Table 16 (Olmo 3 chat evaluation suite) · retrieved 2026-10-15 -
Dataset MATH (Mathematics Aptitude Test of Heuristics) · test portion excluded from fine-tuning data
Evidence
“to filter out the portions of all post-training data (all three stages) that matched the evaluation sets.”
arxiv.org · paper · Section 4.2.1, Step 4 · retrieved 2026-10-15 “We used n-gram matching with 8-grams and an overlap threshold of 0.5”
arxiv.org · paper · Section 4.2.1, Step 4 · retrieved 2026-10-15 “Minerva MATH (2022)”
arxiv.org · paper · Table 16 (Olmo 3 chat evaluation suite) · 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 (2024-12-01)
Evidence (3)
“Date cutoff:** Dec. 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
Post-training-only release on top of the Olmo 3 bases; the Olmo 3 report (v2) covers Olmo 3.1 results (Tables 14, 25). Exclusions copied from the Olmo 3 post-training decontamination, which the report applies to all post-training data.
Sources
- Olmo 3 blog post (Olmo 3.1 update) · website · retrieved 2026-10-15
- Olmo-3.1-32B-Think card · model card · retrieved 2026-10-15
- Olmo-3.1-32B-Instruct card · model card · retrieved 2026-10-15
- Olmo 3 report · paper · retrieved 2026-10-15