Open Models Almanac
Methodsdata curationdata sourcing

methods · data curation

data sourcing

37 methods filed at this node or below it, from the sources of 22 models.

In this branch 37

Everything filed at this node or below it, with one collapsible heading per child node.

filed here 37

Expert co-created training data used · 2 sources · 2 quotes
GitHub Crawl used · 2 sources · 2 quotes
34 single mentions
Bilingual German-English data curation core · 1 source · 1 quote
Conversion to WebDataset tar shards core · 1 source · 1 quote
EgoScale core · 1 source · 1 quote
Pre-training on sensorized human video core · 1 source · 1 quote
Real-world robot-data pre-training core · 1 source · 1 quote
Common Crawl used · 1 source · 1 quote
Community-driven data used · 1 source · 1 quote
DAgger-style data collection used · 1 source · 1 quote
EssentialWeb used · 1 source · 1 quote
Evenly spaced video-frame sampling used · 1 source · 1 quote
FinePDFs used · 1 source · 1 quote
High-recall web-data curation used · 1 source · 1 quote
Long-context training-data construction used · 1 source · 1 quote
Long-document curation used · 1 source · 1 quote
Nemotron-3-Ultra corpus used · 1 source · 1 quote
Nemotron-CC used · 1 source · 1 quote
Nemotron-Post-Training-v3 used · 1 source · 1 quote
Pre-training on Open X-Embodiment data used · 1 source · 1 quote
Resampling training data used · 1 source · 1 quote
Video frame extraction at 1 FPS used · 1 source · 1 quote
Web-data training used · 1 source · 1 quote
Multimodal processor video-sampling parameters optional · 1 source · 1 quote
Internal offline web mentioned · 1 source · 1 quote

By model

Which of this branch's methods each model's own documents describe, and how strongly. Under each model: its strongest treatment anywhere in the branch.

Modelmethods
DeepSeek-V4.1-Flash usedStar-threshold filtering of public GitHub repositories usedVoluntary workflow interaction-data collection used—
Hy4-preview usedExpert co-created training data used—
NVIDIA-Nemotron-3-Ultra-550B-A55B usedCommon Crawl usedEssentialWeb usedFinePDFs usedGitHub Crawl usedNemotron-3-Ultra corpus usedNemotron-CC usedNemotron-Post-Training-v3 used—
DeepSeek-V4-Pro usedLong-document curation used—
FLUX 3 Action usedVideo-game and first-person human-hand footage used—
GLM-5.2 usedWeb knowledge graph construction and question generation used—
GR00T N1.7 coreEgoScale corePre-training on sensorized human video core—
Gemma 4 31B usedVideo frame extraction at 1 FPS used—
Kimi K3 usedAgent-driven recursive knowledge graph construction used—
Kolibri-1 coreBilingual German-English data curation coreThird-party data-source compliance screening coreLong-context training-data construction usedThird-party data-source provenance screening used—
LFM3-1.7B usedMultilingual vision-language training through translated data used—
Laguna-S-2.1 usedHigh-recall web-data curation usedStreaming training-data ingestion with prefetched shared-memory batches used—
LingBot-VLA 2.0 coreReal-world robot-data pre-training coreLarge-scale multi-source pre-training data curation used—
Mistral-Large-4 mentionedInternal offline web mentioned—
MolmoAct 2 usedEvenly spaced video-frame sampling usedSeparate collection of atomic motion primitives used—
NVIDIA-Nemotron-3.5-Lightning-30B-A3B usedGitHub Crawl used—
OpenVLA / OpenVLA-OFT usedPre-training on Open X-Embodiment data used—
Qwen3.6-35B-A3B optionalMultimodal processor video-sampling parameters optional—
SmolVLA corePre-training on community-contributed robotics datasets coreCommunity-driven data used—
VLA Foundry (Foundry-VLA-1.7B) coreConversion to WebDataset tar shards coreResampling training data used—
X-VLA usedDAgger-style data collection used—
π0.5 usedWeb-data training used—
↑