Hugging Face's logo

Collections

Discover the best community collections!

Collections including paper arxiv:2610.19343
Thaki Speech — Korean Voice AI (ASR · TTS · Voice Command)
Korean voice AI: ASR (5 dialects, Nemotron LoRA), VoxCPM2 TTS, spoken-command understanding, voice persona. Pre-registered gates.
Ko-Voice-Command — Korean Spoken Command Understanding
Small readers on a frozen EmbeddingGemma 2 audio tower: Korean voice command → intent, action, entity, decision. No execution.
ThakiCloud Security Models
Security-domain models from ThakiCloud. Experimental previews state their preregistered verdicts and limits on each card.
Satoori-KO — Korean Regional Dialect 27B
Qwen3.8-27B aligned to five Korean regional dialects. Real-data vs. synthetic arm on one benchmark. Not a chat model — single-turn only.
Human-KO — Korean Style-Aligned 27B
Full-weight checkpoints aligning Qwen3.8-27B to Korean conversational style: style, CJK suppression, safety DPO, serving quantization.
Thaki-PolicyGate — Korean regulation-condition checking
9B extractor for Korean case + regulation text, plus a rule backend that verifies its JSON IR and abstains when unsure. Specialist model.
Thaki Speech — Korean Voice AI (ASR · TTS · Voice Command)
Korean voice AI: ASR (5 dialects, Nemotron LoRA), VoxCPM2 TTS, spoken-command understanding, voice persona. Pre-registered gates.
Ko-Voice-Command — Korean Spoken Command Understanding
Small readers on a frozen EmbeddingGemma 2 audio tower: Korean voice command → intent, action, entity, decision. No execution.
Thaki-PolicyGate — Korean regulation-condition checking
9B extractor for Korean case + regulation text, plus a rule backend that verifies its JSON IR and abstains when unsure. Specialist model.
ThakiCloud Security Models
Security-domain models from ThakiCloud. Experimental previews state their preregistered verdicts and limits on each card.
Satoori-KO — Korean Regional Dialect 27B
Qwen3.8-27B aligned to five Korean regional dialects. Real-data vs. synthetic arm on one benchmark. Not a chat model — single-turn only.
Human-KO — Korean Style-Aligned 27B
Full-weight checkpoints aligning Qwen3.8-27B to Korean conversational style: style, CJK suppression, safety DPO, serving quantization.