datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
swe-rebench-v2-CodeWorldModeling
SWE-rebench V2 — CodeWorldModeling Traces
This is a derived dataset. Every record is produced from an instance of
nebius/SWE-rebench-V2.
It is governed by the SWE-rebench V2 license — see License
below — including the requirement to respect each source repository's own
license.
Line-by-line Python execution traces for the test suites of SWE-rebench V2
instances, captured by running each instance's tests under a tracer inside
Nebius ConTree sandboxes.
Each instance comes with a fix… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/swe-rebench-v2-CodeWorldModeling.grug-moe-mix-swarm
Grug-MoE Data-Mix Experiments
The default config contains the original 840-run Fisher-DSP swarm. The harrier_18t75_d768 config contains the Harrier experiments described below.
Fisher-DSP swarm (default)
840 MoE pretraining runs from the Grug-MoE Fisher-DSP data-mixing swarm (d512, TPU / us-central2).
Each run trains on a distinct data mixture over 168 datakit buckets; the swarm is used to regress
mixture weights → eval loss and predict an optimized pretraining… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/grug-moe-mix-swarm.token-counts
Marin Token Counts
Token counts for all datasets used in Marin pretraining runs.
Schema
Column
Type
Description
dataset
string
Dataset identifier
marin_tokens
int
Number of tokens after tokenization
category
string
Content domain (web, code, math, academic, books, etc.)
synthetic
bool
Whether the data is LLM-generated or LLM-translated
Categories
web — Quality-classified Common Crawl text (Nemotron-CC)
code — Source code and… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/token-counts.marinfold-exp11-protein-docs-seq
marinfold-exp11-pdocs-seq
Sequence-only derivative of
eczech/marinfold-exp11-protein-docs.
For every row, the document field has been reduced to just the amino-acid sequence
portion: the <begin_sequence> tag followed by the per-residue three-letter tokens
(e.g. <begin_sequence> <MET> <LYS> <ASN> ...). The <contacts-and-distances-v1>
document-type prefix and everything from <begin_statements> onward (contacts and
distances) are removed. The token format is preserved verbatim so… See the full description on the dataset page: https://huggingface.co/datasets/eczech/marinfold-exp11-protein-docs-seq.mcp-atlas-easy
MCP-Atlas-Easy
An easy, single-tool-call benchmark for pretrained (base) language models, derived from ScaleAI/MCP-Atlas.
MCP-Atlas evaluates instruction-tuned agents on multi-step tool orchestration (3–6 calls per task across 36 real MCP servers). MCP-Atlas-Easy strips that down to the simplest possible form of the same skill: one tool spec, one trivially unambiguous request, one correct tool call, then stop. This makes it usable as a completion-style eval for base models with… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/mcp-atlas-easy.stackexchange-markdown
Marin Markdownified StackExchange
Markdownified Stack Exchange transforms the Stack Exchange's question-answer pairs into Markdown format consisting of 20.4B tokens. This dataset preserves the content contained in technical discussions while organizing it into a thread format for language model training.
Value
Tokens
20 413 785 853
Primary source
https://archive.org/details/stackexchange
File format
JSONL
License
CC (mirrors upstream SE licenses)… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/stackexchange-markdown.ar5iv-no-problem-markdown
Marin Markdownified Ar5iv
Markdownified Ar5iv transforms academic papers from arXiv into clean, structured Markdown format consisting of 2.74B tokens across two splits. This dataset preserves th content while making it accessible for language model training on academic text.
Value
Tokens
2 742 463 924
Primary source
https://sigmathling.kwarc.info/resources/ar5iv-dataset-2024/
File format
JSONL
License
C-UDA-1.0 (mirrors upstream Ar5iv licenses)… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/ar5iv-no-problem-markdown.marinfold-exp11-protein-docs
marinfold-exp11-pdocs
Quality-bucketed re-publication of the contacts-and-distances-v1-5x config from
timodonnell/protein-docs,
partitioned by the source round column:
Config
Source rounds
Approx rows
high
round 0
~1.68M
medium
round 1
~1.42M
low
round 2–4
~2.29M
Train/val/test split assignment is inherited from the source dataset (leakage-resistant
structural-cluster hashing). All columns from the source are preserved; rows are simply
partitioned by round.
See… See the full description on the dataset page: https://huggingface.co/datasets/eczech/marinfold-exp11-protein-docs.openthoughts4-code-9168-prompts-qwen3-30b-a3b-thinking-2507-n16-flattened-logprobs-k16
OpenThoughts-4 Code SDG: Qwen3-30B-A3B-Thinking-2507 (n=16, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-30B-A3B-Thinking-2507
on the Marin OpenThoughts-4 code SDG prompt
set.
Each prompt is sampled n=16 times, and for every generated token the dataset
stores the chosen-token log probability plus the top-16 log probabilities
over the vocabulary, enabling distillation, KL-style fine-tuning,
reranking, and uncertainty analysis.
Generation setup
Field… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-code-9168-prompts-qwen3-30b-a3b-thinking-2507-n16-flattened-logprobs-k16.openthoughts4-code-9168-prompts-qwen3-32b-n16-flattened-logprobs-k16
OpenThoughts-4 Code SDG: Qwen3-32B (n=16, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-32B
on the Marin OpenThoughts-4 code SDG prompt
set.
Each prompt is sampled n=16 times, and for every generated token the dataset
stores the chosen-token log probability plus the top-16 log probabilities
over the vocabulary, enabling distillation, KL-style fine-tuning,
reranking, and uncertainty analysis.
Generation setup
Field
Value
Generator model… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-code-9168-prompts-qwen3-32b-n16-flattened-logprobs-k16.ar5iv-warning-markdown
Marin Markdownified Ar5iv
Markdownified Ar5iv transforms academic papers from arXiv into clean, structured Markdown format consisting of 22.34B tokens across two splits. This dataset preserves th content while making it accessible for language model training on academic text.
Value
Tokens
19 552 307 274
Primary source
https://sigmathling.kwarc.info/resources/ar5iv-dataset-2024/
File format
JSONL
License
C-UDA-1.0 (mirrors upstream Ar5iv licenses)… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/ar5iv-warning-markdown.openthoughts4-code-9168-prompts-glm-5.2-n4
OpenThoughts-4 Code — GLM-5.2 n=4
Quality-filtered synthetic responses from
zai-org/GLM-5.2-FP8 for the
9,168 unique instruction_seed values in
mlfoundations-dev/hero_run_4_code.
Each prompt has four accepted responses, for 36,672 rows total.
Generation
Field
Value
Generator
zai-org/GLM-5.2-FP8
Samples per prompt
4
Temperature
1.0
Top-p
0.95
Maximum generated tokens
256,000
Thinking mode
enabled
Inference engine
vLLM on 8 GB200 GPUs… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-code-9168-prompts-glm-5.2-n4.openthoughts4-science-26041-prompts-qwen3-30b-a3B-thinking-2507-n8-flattened-logprobs-k16
OpenThoughts-4 Science SDG: Qwen3-30B-A3B-Thinking-2507 (n=8, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-30B-A3B-Thinking-2507
on the Marin OpenThoughts-4 science SDG prompt
set.
Each prompt is sampled n=8 times, and for every generated token the dataset
stores the chosen-token log probability plus the top-16 log probabilities
over the vocabulary, enabling distillation, KL-style fine-tuning,
reranking, and uncertainty analysis.
Generation setup
Field… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-science-26041-prompts-qwen3-30b-a3B-thinking-2507-n8-flattened-logprobs-k16.wikipedia-markdown
Marin Markdownified Wikipedia
Markdownified Wikipedia is a large-scale, pre-processed version of the English Wikipedia Enterprise HTML dump consisting of 8.59B tokens. The corpus has been converted to clean, section-aware Markdown for language-model training.
Value
Tokens
8 587 224 558
Primary source
https://dumps.wikimedia.org/other/enterprise_html/runs/20241201/enwiki-NS0-20241201-ENTERPRISE-HTML.json.tar.gz
File format
JSONL
License
CC-BY-SA 4.0 (mirrors… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/wikipedia-markdown.openthoughts4-science-26041-prompts-qwen3-32b-n8-flattened-logprobs-k16
OpenThoughts-4 Science SDG: Qwen3-32B (n=8, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-32B
on the Marin OpenThoughts-4 science SDG prompt
set.
Each prompt is sampled n=8 times, and for every generated token the dataset
stores the chosen-token log probability plus the top-16 log probabilities
over the vocabulary, enabling distillation, KL-style fine-tuning,
reranking, and uncertainty analysis.
Generation setup
Field
Value
Generator model… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-science-26041-prompts-qwen3-32b-n8-flattened-logprobs-k16.identity-data
Identity Data
English synthetic conversations for reinforcing model identity and provenance.
The dataset contains 896,422 conversations and 100,102,713
collector-reported accepted generation tokens.
Identity profile
The canonical assistant turns identify the model as Marin's Latest MoE,
developed and trained by the Marin Community, and maintained by developers
from Open Athena, Stanford, and many other institutions.
Some conversations acknowledge contributions… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/identity-data.glm-5.2-kernelgym-rollouts
GLM-5.2 KernelGym Rollouts
This dataset contains 3,200 feedback-driven GPU-kernel optimization trajectories
generated by zai-org/GLM-5.2-FP8: 100 validation tasks, two backends (inline
CUDA and Triton), and 16 rollouts per task.
Each trajectory retains the prompt/feedback message history, model responses and
reasoning, extracted kernel code, KernelGym compilation and correctness results,
profiling metadata, token usage, and stopping decision. Every published record
ended with… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/glm-5.2-kernelgym-rollouts.openthoughts4-code-9168-prompts-qwen3-4b-n16-flattened-logprobs-k16
OpenThoughts-4 Code SDG: Qwen3-4B (n=16, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-4B
on the Marin OpenThoughts-4 code SDG prompt
set.
Each prompt is sampled n=16 times, and for every generated token the dataset
stores the chosen-token log probability plus the top-16 log probabilities
over the vocabulary, enabling distillation, KL-style fine-tuning,
reranking, and uncertainty analysis.
Generation setup
Field
Value
Generator model
Qwen/Qwen3-4B… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-code-9168-prompts-qwen3-4b-n16-flattened-logprobs-k16.open-thoughts-4-30k-code-qwen3-32b-annotated-32768-tokens
Dataset Card for Open-Thoughts-4-30K-Code-Qwen3-32B-Annotated-32768-Tokens
Overview
This dataset is a variant of marin-community/open-thoughts-4-30k-code-qwen3-32b-annotated with an extended maximum sequence length. The responses in the generated_text column were generated with max output tokens = 32768 (instead of 7500 in the original dataset), allowing for longer and more complete chain-of-thought reasoning.
Generation Details
Model: Qwen/Qwen3-32B… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/open-thoughts-4-30k-code-qwen3-32b-annotated-32768-tokens.open-thoughts-4-11k-math-qwen3-32b-agreed-answers
OpenThoughts4 11K Math - Qwen3-32B Agreed Answers
Overview
This dataset contains 10,953 samples from the OpenThoughts4 math dataset where both Qwen3-32B and Qwen3-235B-A22B produced valid \boxed{} answers that match exactly. This dataset contains the Qwen3-32B reasoning traces.
Relationship to Other Datasets
This is one of 10 child datasets derived from two parent datasets:
Parent datasets (29,963 samples each):… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/open-thoughts-4-11k-math-qwen3-32b-agreed-answers.openthoughts4-science-26041-prompts-qwen3-4b-n8-flattened-logprobs-k16
OpenThoughts-4 Science SDG: Qwen3-4B (n=8, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-4B
on the Marin OpenThoughts-4 science SDG prompt
set.
Each prompt is sampled n=8 times, and for every generated token the dataset
stores the chosen-token log probability plus the top-16 log probabilities
over the vocabulary, enabling distillation, KL-style fine-tuning,
reranking, and uncertainty analysis.
Generation setup
Field
Value
Generator model… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-science-26041-prompts-qwen3-4b-n8-flattened-logprobs-k16.question-answer-Subject-Finance-Instructopen-thoughts-4-11k-math-qwen3-235b-a22b-agreed-answers
OpenThoughts4 11K Math - Qwen3-235B-A22B Agreed Answers
Overview
This dataset contains 10,953 samples from the OpenThoughts4 math dataset where both Qwen3-32B and Qwen3-235B-A22B produced valid \boxed{} answers that match exactly. This dataset contains the Qwen3-235B-A22B reasoning traces.
Relationship to Other Datasets
This is one of 10 child datasets derived from two parent datasets:
Parent datasets (29,963 samples each):… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/open-thoughts-4-11k-math-qwen3-235b-a22b-agreed-answers.open-thoughts-4-6k-math-qwen3-235b-a22b-disagreed-answers
OpenThoughts4 6K Math - Qwen3-235B-A22B Disagreed Answers
Overview
This dataset contains 5,982 samples from the OpenThoughts4 math dataset where both Qwen3-32B and Qwen3-235B-A22B produced valid \boxed{} answers that differ. This dataset contains the Qwen3-235B-A22B reasoning traces.
Relationship to Other Datasets
This is one of 10 child datasets derived from two parent datasets:
Parent datasets (29,963 samples each):… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/open-thoughts-4-6k-math-qwen3-235b-a22b-disagreed-answers.marin-8b-tb2-eval-traces
Marin-8B Terminal-Bench 2.0 Eval Traces
Per-trial agent trajectories from running Terminal-Bench 2.0 against:
marin_8b_base — marin-community/marin-8b-instruct (base, no SFT). 0 / 89 = 0.0% unique solves (89 canonical trials).
marin_8b_sft_step5720 — exp4420 SFT (Marin-8B Instruct + 366K Nemotron-Terminal-Corpus, 5,721 steps, 2 epochs, 32K context). 1 / 89 = 1.1% unique solves (only modernize-scientific-stack; 89 canonical trials).
Posted alongside the diagnosis in… See the full description on the dataset page: https://huggingface.co/datasets/AlienKevin/marin-8b-tb2-eval-traces.marine-species-zh
海纳海洋生物数据集 | Marine Species Dataset (Chinese)
4662 个海洋物种的结构化数据集：学名、中文俗名、完整分类阶元、命名人、OBIS 分布记录数、图片链接（逐行标注授权协议与作者署名）、中文简介。
A structured dataset of 4,662 marine species: scientific names, Chinese vernacular names, full taxonomy, authorities, OBIS occurrence counts, image links (with per-row license and author attribution), and 2,120 Chinese descriptions translated/organized from Chinese Wikipedia.
数据来自海洋科普公益平台 海纳 · hainahub.cn 的物种图鉴底层，随图鉴扩容滚动更新。
数据规模 / Stats
指标… See the full description on the dataset page: https://huggingface.co/datasets/hainahub/marine-species-zh.minicpm5-vivamais-text-sft-v4
MiniCPM5 Viva Mais text SFT v4
This dataset contains the redacted training and evaluation artifacts used for
marinarosa/minicpm5-1b-vivamais-v4. It was built for Viva Mais, a local-first Portuguese WhatsApp
travel-agency copilot that answers grounded questions from an extracted CRM
context.
Files
data/train.jsonl: 4000 chat-format SFT rows.
data/eval/vivamais_qa_eval.jsonl: 158 dashboard QA eval
rows.
data/teacher/rio31_teacher_distill.jsonl: 80 accepted
rows… See the full description on the dataset page: https://huggingface.co/datasets/marinarosa/minicpm5-vivamais-text-sft-v4.open-thoughts-4-6k-math-qwen3-32b-neither-has-boxed
OpenThoughts4 6K Math - Qwen3-32B Neither Has Boxed Answer
Overview
This dataset contains 6,110 samples from the OpenThoughts4 math dataset where neither Qwen3-32B nor Qwen3-235B-A22B produced a valid \boxed{} answer. This dataset contains the Qwen3-32B reasoning traces.
Relationship to Other Datasets
This is one of 10 child datasets derived from two parent datasets:
Parent datasets (29,963 samples each):… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/open-thoughts-4-6k-math-qwen3-32b-neither-has-boxed.open-thoughts-4-5k-math-qwen3-32b-only-235b-has-boxed
OpenThoughts4 5K Math - Qwen3-32B Only 235B Has Boxed Answer
Overview
This dataset contains 5,441 samples from the OpenThoughts4 math dataset where only Qwen3-235B-A22B produced a valid \boxed{} answer (Qwen3-32B did not). This dataset contains the Qwen3-32B reasoning traces (which lack valid boxed answers).
Relationship to Other Datasets
This is one of 10 child datasets derived from two parent datasets:
Parent datasets (29,963 samples each):… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/open-thoughts-4-5k-math-qwen3-32b-only-235b-has-boxed.minicpm5-vivamais-text-sft-v1
MiniCPM5 Viva Mais Text SFT v1
This dataset is the exact JSONL training/evaluation package used for the
MiniCPM5 Viva Mais text QA candidate v1 run.
Files
minicpm5_text_sft.jsonl: 12,000 SFT rows.
vivamais_qa_eval.jsonl: 32 fixed Viva Mais dashboard QA eval rows.
Training Mix
The SFT mix was generated by the Viva Mais repository pipeline from the Modal
volume minicpm5-vivamais-text-data:
2,400 rows from Polygl0t/gigaverbo-v2-sft
5,400 Viva Mais… See the full description on the dataset page: https://huggingface.co/datasets/marinarosa/minicpm5-vivamais-text-sft-v1.