Open Models Almanac
Methodsdata curationsynthetic data

methods · data curation

synthetic data

138 methods filed at this node or below it, from the sources of 24 models.

In this branch 138

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

filed here 138

Large-scale agentic task synthesis pipeline core · 3 sources · 3 quotes
Synthetic data generation used · 3 sources · 3 quotes
Counterfactual data augmentation mentioned · 2 sources · 2 quotes
128 single mentions
Gaussian-noise augmentation of depth maps used · 1 source · 1 quote
Knowledge-graph-guided task synthesis used · 1 source · 1 quote
Metadata-conditioned synthetic generation used · 1 source · 1 quote
Minimal-tool autonomous task authoring used · 1 source · 1 quote
Regenerate training-data completions used · 1 source · 2 quotes
Rubric-guided RL data used · 1 source · 1 quote
Synthetic English pre-training rephrasing used · 1 source · 1 quote
Synthetic environment generation pipelines used · 1 source · 1 quote
Synthetic German web-data rephrasing used · 1 source · 1 quote
Synthetic privacy-preserving personas used · 1 source · 1 quote
Task-specific data injection used · 1 source · 1 quote
Template-based synthetic data generation used · 1 source · 1 quote
User-simulated conversational continuation used · 1 source · 3 quotes
VLM-based offline automatic CoC labeling used · 1 source · 1 quote
Co-training with generated neural trajectories evaluated · 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 coreAutomated large-scale data synthesis and environment construction coreAutomated synthesis of agent tasks and environments with progressive scaling coreProcedural construction and scaling of interactive agent environments coreSynthesis of verifiable training tasks, reference solutions, and reward signals coreAgent screening of coding projects for container execution and automatic verification usedAgent-designed coding tasks with fail-to-pass and pass-to-pass evaluation points usedIterative reward-guided training for task construction used—
Hy4-preview usedTraining-data construction around expert work products used—
GLM-5.3 usedSynthetic environment generation pipelines usedTraining with vulnerability-focused data and environments usedVerifier synthesis without reference solution used—
MiMo-V2.6-Flash usedAutomated large-scale data synthesis and environment construction usedMulti-agent synthesis of local software mocks for reproducible environments usedTraining SFT teachers on synthetic demonstrations for open-domain tasks used—
NVIDIA-Nemotron-3-Ultra-550B-A55B usedKnowledge distillation for synthetic data generation usedSynthetic data generation usedCounterfactual data augmentation mentioned—
Alpamayo 1.5 / 2 Super coreCausally grounded, decision-linked reasoning traces coreTemporal separation of historical and future inputs in CoC auto-labeling usedVLM-based automatic labeling with heuristic rules usedVLM-based offline automatic CoC labeling used—
DeepSeek-V3.2 coreAutomatic synthesis of task-oriented RL environments coreLarge-scale agentic task synthesis pipeline coreAutomated environment, toolset, task, and solution synthesis pipeline usedCode-agent environment construction from GitHub issue-PR pairs usedHeterogeneous answer-generation agents producing diverse candidate responses usedIterative task-difficulty escalation with solution and verifier updates usedMulti-agent pipeline for search-agent training data generation usedQuestion-construction agent that explores entities with search tools usedSearch-capable verification agent for answer validation used—
DeepSeek-V4-Pro usedRubric-guided RL data usedTask-specific data injection used—
GLM-5.2 usedAgent-based synthetic-data generation pipeline used—
GR00T N1.7 usedInverse-dynamics-model pseudo-action labeling usedNeural trajectory generation using video generative models usedOWL-v2 target-object bounding-box annotation usedRecaptioning videos that fail instruction-adherence review usedSimulation trajectory generation through demonstration transformation and replay usedCo-training with generated neural trajectories evaluated—
Inkling usedRandomized tool set and schema during training usedSynthetic data generation used—
Kimi K3 usedAgentic visual reasoning with a Python interpreter in a sandbox usedKnowledge and mathematics corpus rephrasing with fidelity verification usedKnowledge-graph-guided task synthesis usedMulti-stage verification with human-in-the-loop annotation usedProgrammatic multimodal data coupling code with rendered visuals usedSynthetic long-context data via permutation and concatenation of multimodal documents usedSynthetic trajectory generation with domain-specialized models used—
Kolibri-1 usedAbstention training with explicit negative examples usedAdd synthesized long-context conversations to the SFT mix usedBackfilling missing reasoning between tool calls usedConcatenate tool-use conversations into long-context synthesis tasks usedEnglish-to-German translation for data augmentation usedGenerate fully German reasoning-training samples usedGenerate unanswerable examples by masking context with helpful and adversarial agents usedGerman document rephrasing for data augmentation usedGerman reasoning prefill for synthetic teacher traces usedMerlin-Arthur synthetic negative-example generation procedure usedMulti-hop retrieval task synthesis using GLM-5.2 usedReference-blind German rewriting of math and science problems usedRegenerate training-data completions usedRegeneration of reasoning traces for agentic SFT data usedRetain both reasoning and non-reasoning trajectory variants usedSearch-dependence filtering for multi-hop questions usedSection-grounded simulated requests for long-context training usedSeeded planning of long-context question locations and answer shape usedSpread long-context queries across document sections usedSynthetic data generation usedSynthetic English pre-training rephrasing usedSynthetic German web-data rephrasing usedSynthetic German-language data for supervised fine-tuning usedSynthetic multi-hop retrieval-question generation usedSynthetic-data training environments for domain-specific skills usedTemplate-based synthetic data generation usedTrain value judgments using everyday requests without explicit ethical cues usedTrain with full-context, Merlin-masked, and Morgana-masked question variants usedUser-simulated conversational continuation usedVerifiable terminal-task generation by injecting repository bugs used—
LFM3-1.7B usedConversion of text SFT datasets into synthetic audio conversations usedDataset recaptioning for multimodal training usedOn-policy preference-data generation and skill-specific refinement usedSynthetic multi-turn audio conversation data generation usedSynthetic paralinguistic and discourse feature generation usedSynthetic VQA generation targeted at visual understanding usedTheme-conditioned synthetic audio conversation generation usedVisual augmentation for robust document OCR usedVoice cloning to diversify synthetic user voices used—
Laguna-S-2.1 coreDynamic multi-agent synthetic-data generation loop coreModular synthetic-data pipeline composition coreBehavioral-constraint augmentation with synthetic system messages usedLarge-scale targeted synthetic data rephrasing usedMatch synthetic-pipeline complexity to teacher capability usedMetadata-conditioned synthetic generation usedMulti-stage synthetic-data generation cascade usedSynthetic data augmentation of organic pre-training data usedTraining-task generation from real commit history usedTwo-sided test validation for synthetic coding tasks used—
LingBot-VLA 2.0 coreAutomated vision-language-model annotation of manipulation videos coreVLM-based temporal subtask segmentation and instruction generation coreGaussian-noise augmentation of depth maps usedGenerate pseudo-depth labels from stereo infrared pairs usedSimulate real active RGB-D camera imaging to generate imperfect depth usedSynthetic RGB-D data generation using self-hosted 3D assets used—
MiMo-V2.6-Pro usedAutomated large-scale data synthesis and environment construction usedMulti-agent synthesis of local software mocks for reproducible environments usedTraining SFT teachers on synthetic demonstrations for open-domain tasks used—
Mistral-Large-4 coreAgent-based synthetic-data generation pipeline coreMinimal-tool autonomous task authoring usedRefine software tasks using multiple solvers and trajectory inspection usedReuse successful authoring-agent trajectories as midtraining data usedSeed terminal-agent tasks with technical content from pretraining data usedTaxonomy-guided generation of simulated workplace environments usedTwo-stage GitHub-derived software task generation and RL refinement used—
MolmoAct 2 usedDemonstration collection with alternative language-conditioned motion paths usedRandomize requested instruction length during VLM relabeling usedVLM-based point annotation for visual traces usedVLM-based robot demonstration language re-annotation used—
NVIDIA-Nemotron-3.5-Lightning-30B-A3B usedAgentic visual tool use for image operations usedBodhan-MT translation with Gemma-4 quality assessment and language-specific thresholds usedChunked synthetic long-context translation data usedConcurrent scheduling of translation and quality checks usedCoverage-audit-driven agentic question generation usedDocument-grounded synthetic question generation with answerability verification usedFour-stage Indic synthetic-data generation pipeline usedKnowledge distillation for synthetic data generation usedLarge-scale synthetic data generation, filtering, and curation usedOverlapping windows for longer Santali documents usedRewrite low-scoring translations using source text and identified errors, then recheck usedScore translations against their source for quality usedStep-structured, visually grounded reasoning traces usedSynthetic graph-structured long-context reasoning data usedSynthetic Indic training-data generation by translating curated English documents usedSynthetic privacy-preserving personas usedSynthetic transduction-task training for long translation usedTranslate each selected source document in full without splitting usedCounterfactual data augmentation mentioned—
NVIDIA-Nemotron-3.5-Super-VL-120B-A12B usedAgentic visual tool use for image operations usedBodhan-MT translation with Gemma-4 quality assessment and language-specific thresholds usedChunked synthetic long-context translation data usedConcurrent scheduling of translation and quality checks usedCoverage-audit-driven agentic question generation usedDocument-grounded synthetic question generation with answerability verification usedFiltering synthetic trajectories for correctness and validity usedFour-stage Indic synthetic-data generation pipeline usedOverlapping windows for longer Santali documents usedRewrite low-scoring translations using source text and identified errors, then recheck usedScore translations against their source for quality usedStep-structured, visually grounded reasoning traces usedSynthetic data generation with NeMo Data Designer usedSynthetic graph-structured long-context reasoning data usedSynthetic Indic training-data generation by translating curated English documents usedSynthetic transduction-task training for long translation usedTranslate each selected source document in full without splitting used—
SmolVLA usedFrame-and-instruction prompting for automatic task annotation usedRewrite task instructions under 30 characters beginning with an action verb usedStandardized action-oriented task descriptions from frames and labels used—
Xiaomi-Robotics-1 coreAutomated language annotation of manipulation trajectories coreCross-validated consensus annotation for visual grounding usedLanguage annotation of human-segmented post-training clips usedPre-trained VLM relabeling to refine VQA and caption annotations usedVLM-generated embodied reasoning and planning data from robot trajectories used—
π0.5 usedExpert-provided language demonstrations of high-level subtasks used—
↑