HF Spaces for demoing chat completion models—no ZeroGPU, WebGPU, or BYOK included. Thank you so much to these devs!
Urro PRO
urroxyz
AI & ML interests
computational linguistics major
Recent Activity
updated a collection about 5 hours ago
WTF GENIUS PAPERS commentedon a paper about 5 hours ago
Omnilingual MT: Machine Translation for 1,600 Languages upvoted a paper about 5 hours ago
Omnilingual MT: Machine Translation for 1,600 LanguagesOrganizations
Tiny (<30B) models that tend to outperform their same-parameter counterparts.
Datasets written by humans and/or reverse-engineered from text with deterministic algorithms. No illegal scraping or unethical synthesis *...mostly.
Papers that made me appreciate my major and my life a little more. obs=Observation, innov=Innovation. Most papers are abt improving tiny models.
-
Diffusion Language Models Know the Answer Before Decoding
Paper • 2508.19982 • Published • 27 -
ReFusion: A Diffusion Large Language Model with Parallel Autoregressive Decoding
Paper • 2512.13586 • Published • 94 -
LSRIF: Logic-Structured Reinforcement Learning for Instruction Following
Paper • 2601.06431 • Published • 12 -
Distribution-Aligned Sequence Distillation for Superior Long-CoT Reasoning
Paper • 2601.09088 • Published • 63
Depending on your definitions, these models may not be strictly "ethical" or "legal", yet they are 100% more ethical and legal than GPT or Claude.
ONNX conversions of ASR models with attentions enabled for output. Especially useful for word-level timestamp extraction.
HF Spaces for demoing chat completion models—no ZeroGPU, WebGPU, or BYOK included. Thank you so much to these devs!
Papers that made me appreciate my major and my life a little more. obs=Observation, innov=Innovation. Most papers are abt improving tiny models.
-
Diffusion Language Models Know the Answer Before Decoding
Paper • 2508.19982 • Published • 27 -
ReFusion: A Diffusion Large Language Model with Parallel Autoregressive Decoding
Paper • 2512.13586 • Published • 94 -
LSRIF: Logic-Structured Reinforcement Learning for Instruction Following
Paper • 2601.06431 • Published • 12 -
Distribution-Aligned Sequence Distillation for Superior Long-CoT Reasoning
Paper • 2601.09088 • Published • 63
Tiny (<30B) models that tend to outperform their same-parameter counterparts.
Depending on your definitions, these models may not be strictly "ethical" or "legal", yet they are 100% more ethical and legal than GPT or Claude.
Datasets written by humans and/or reverse-engineered from text with deterministic algorithms. No illegal scraping or unethical synthesis *...mostly.
ONNX conversions of ASR models with attentions enabled for output. Especially useful for word-level timestamp extraction.