AI Train2026
End-to-end Python framework for training and serving NVIDIA's Cosmos world models (Cosmos3), integrating distributed training (FSDP/TP/CP/PP), DCP/safetensors checkpoints, dataset adapters, multiple inference backends, online serving, and agent skills.
AI Dataset2026
Supervised fine-tuning dataset of instruction-style examples in English and Chinese covering generation, QA, reasoning, math and code — targeted for SFT of 10–100B-parameter LLMs. Associated with arXiv:2602.09003; first published May 21, 2026.
AI Model2026
Fine-tuned reasoning model that speeds up structured multi-step outputs using Multi-Token Prediction (MTP) from a Qwen3.6-27B base. Produces more concise, faster generations for coding, DevOps, math, and constrained-format tasks; experimental community release for research and evaluation.
AI Dataset2026
Provides 100,000 generated low-quality↔high-quality image pairs created with modern multi-frame/multi-modal models to boost generalization of image restoration methods; includes train/test JSONL lists, baseline training code, and pretrained checkpoints under CC BY‑NC‑ND 4.0.
AI Dataset2026
Provides raw newline-delimited JSON agent traces where assistant responses were generated by qwen/qwen3.7-max, captured with Teich; includes 47 JSONL files, an embedded tools schema snapshot, and conversion guidance for supervised fine‑tuning and distillation.
AI Model2026
Performs hour-scale video understanding and fine-grained temporal localization while exposing agent-style multimodal tool/code/search abilities. Built on a sparse-attention long-context architecture (DSA) and a specialized inference stack—best used in GPU-backed research or production evaluation.
AI Dataset2026
Metadata-only corpus of 146.3M new GitHub source-code files (commit_id, rel_path, language) intended as an incremental update to Nemotron v1/v2 for LLM code pretraining; CC-BY-4.0 licensed and designed to be used jointly with older versions.
Introduces Draft-OPD, an on-policy distillation method for training lightweight draft models used in speculative decoding — it focuses learning on draft-induced errors via target-assisted rollouts and replay, improving acceptance length and enabling >5× lossless LLM inference acceleration.
Proposes TrOPD, a method that restricts token-level on-policy distillation to regions where teacher supervision is reliable to stabilize training under teacher–student distribution mismatch. Adds outlier handling (clipping, masking, forward-KL) and off-policy guidance; shows consistent gains on math reasoning, code generation and general benchmarks.
Studies small trainable adapters (PEFT) used as persistent personal models on top of large foundation models, analyzing three scaling axes—Scale Up, Scale Down, Scale Out—and introducing MinT, an infrastructure for adapter identity, provenance, evaluation, and serving.
Trains a GPT-style causal Transformer on a 2-billion-frame retargeted motion corpus to enable zero-shot whole-body motion tracking and control. By scaling both data and model capacity, it tracks highly dynamic behaviors while generalizing to unseen motions; accepted to CVPR 2026.
Explores how training recipe — data composition, teacher guidance, and task mixture — shapes few-step distillation for text-to-image generation and instruction-guided image editing; introduces Qwen-Image-Flash and empirical findings that training pipeline organization matters as much as distillation objectives.