Tag

gemma-4-31B-it-DFlash

Draft model for speculative decoding that uses a lightweight block-diffusion drafter to propose multiple tokens in parallel; designed to pair with google/gemma-4-31B-it and accelerate autoregressive text generation (official benchmarks report up to ~5.8× throughput).

L2P: Unlocking Latent Potential for Pixel Generation

Transforms pretrained latent-diffusion priors into pixel-space diffusion models by removing the VAE and training shallow pixel layers on LDM-generated synthetic images — enabling fast convergence, native 4K output, and low-data training on 8 GPUs.

stanford-vision-lab/gpic

Provides paired images and English captions for vision–language research, curated by Stanford Vision Lab and hosted on Hugging Face; useful for training and evaluating multimodal models and reproducing related research.

LOCUS v1.0

Provides a county-harmonized corpus of U.S. municipal and county ordinance text (≈2.21M chunks) labeled for function, substantive indicator, and topic to support legal NLP, retrieval, and comparative local-law research. Includes model-assigned labels and continuous scorers (opacity, paternalism, enforcement discretion) plus coverage metadata; not exhaustive or a substitute for legal advice.

Preview of an MoE model family (V4-Pro: 1.6T params, 49B active; V4-Flash: 284B, 13B active) built for 1M-token contexts. A hybrid attention design cuts single-token inference FLOPs to 27% and KV cache to 10% versus V3.2 at million-token length.

IndustryBench

Multilingual benchmark for evaluating LLMs' industrial domain knowledge via 2,049 expert-curated QA pairs spanning 10 product verticals and four languages, with each item grounded to industry or national standards and an LLM-as-judge evaluation pipeline.

L2P: Unlocking Latent Potential for Pixel Generation

Transfers pretrained latent diffusion priors into pixel space to train pixel-space diffusion models using only synthetic images from LDMs. Trains shallow pixel layers while freezing most LDM internals, reducing data and compute needs and enabling native 4K generation without a VAE.

GoLongRL (Kwai-Klear)

RL training dataset for long-context language-model fine-tuning with ~23K samples and nine reward types, provided in Parquet with bilingual ground-truth and reward metadata for direct RL/bench evaluation.

GGT-100K: Generative Ground Truth for Generalizable Real-World Image Restoration

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.

Qwen-Image-Bench

Creator-centric benchmark for evaluating text-to-image models with 1,000 bilingual prompts and a 3-level, 56-facet taxonomy. Includes a trained Q-Judger judge model and leaderboard-ready evaluation scripts to surface gaps in real-world fidelity and creative generation.

WBench

Provides a 289-case (1,058-turn) multi-turn benchmark that evaluates interactive video world models across 22 metrics and five dimensions (quality, setting, interaction, consistency, physics). Includes first-/third-person and navigation splits plus a 20-model leaderboard for head-to-head comparisons.

Skill0.5: Joint Skill Internalization and Utilization for Out-of-Distribution Generalization in Agentic Reinforcement Learning

Combines internalizing general skills with task-specific skill utilization via a difficulty-aware router to improve in-distribution and out-of-distribution performance for agentic RL. Uses privileged distillation for hard tasks and diagnostic probing for easy tasks; evaluated on ALFWorld and WebShop.