AI & ML interests
None defined yet.
Recent Activity
View all activity
Organization Card
Welcome to Lunark AI 👋
Lunark is a relational AI platform that transforms one-way content into two-way emotional experiences. We build AI systems that understand context, express emotions, and maintain long-term memory—enabling characters to form meaningful relationships with users.
Our flagship product KAIdol is an emotionally immersive AI idol chatbot where fans interact with K-POP trainee characters that remember, grow, and respond with genuine personality.
🌐 lunark.ai | 🎤 kaidol.lunark.ai
Featured Models & Resources
All models, datasets, and demos are published under our development hub:
🎤 KAIdol Character AI Models
Fine-tuned LLMs for emotionally immersive, persona-consistent character conversations. Trained with SFT, DPO, KTO, and knowledge distillation across multiple base architectures.
| Model | Base | Size | Method |
|---|---|---|---|
| kaidol-exaone-7.8b-v7 | EXAONE 3.5 | 7.8B | SFT (Latest) |
| kaidol-qwen2.5-14b-v7 | Qwen2.5 | 14B | SFT (Latest, Large) |
| kaidol-qwen2.5-7b-dpo-v5 | Qwen2.5 | 7B | DPO (Preference-optimized) |
| kaidol-exaone-7.8b-kimi | EXAONE 3.5 | 7.8B | Distillation (Kimi) |
| doha-kto | — | 4B | KTO (Per-character tuning) |
🎭 Korean Roleplay Models
Large-scale Korean roleplay and conversation models for creative, interactive storytelling.
| Model | Base | Size | Description |
|---|---|---|---|
| Qwen3-30B-Korean-Roleplay | Qwen3 | 31B | Korean RP specialized |
| kaidol-qwen3-14b-korean-rp | Qwen3 | 14B | Lightweight RP |
| Qwen3-30B-A3B-Kaidol-v5 | Qwen3 MoE | 31B | KAIdol-optimized |
🛠️ NLP Tools & Utilities
Specialized models supporting the KAIdol conversation pipeline.
| Model | Task | Size | Description |
|---|---|---|---|
| kaidol-ner-multilingual | Token Classification | 0.3B | Multilingual NER |
| kaidol-thinking-sft-4b | Text Generation | 4B | Enhanced reasoning |
| Qwen2-14B-Lunark-No4-v2-GGUF | GGUF | 15B | Quantized for local inference |
📚 Datasets
Open datasets for character AI and Korean roleplay research.
| Dataset | Samples | Description |
|---|---|---|
| kaidol-sft-v2 | 65.1K | SFT training conversations |
| kaidol-roleplay-v2 | 1.49K | Roleplay dialogue data |
| korean-character-roleplay-sft | 1.07K | Korean character RP SFT |
| kaidol-character-dataset | 2.11K | Character profiles & settings |
| kaidol-thinking-sft-dataset | 58.8K | Thinking & reasoning data |
🚀 Demos & Spaces
| Space | Description |
|---|---|
| KAIdol Thinking Experiment | Compare model responses across K-pop idol scenarios (T4 GPU) |
| kaidol-sft-pipeline | Side-by-side SFT response comparison pipeline |
Tech Stack Highlights
| Area | Details |
|---|---|
| Emotion Recognition | 12+ fine-grained emotions with real-time inference |
| Hybrid Memory | Redis + Firestore + Qdrant + Vertex AI Memory Bank |
| Agent Orchestration | Vertex AI ADK + LangChain for memory, emotion & relationship-aware conversations |
| Training Methods | SFT, DPO, KTO, Knowledge Distillation — Qwen, EXAONE, SOLAR, VARCO, Llama |
| Evaluation | LLM-as-Judge with Opik tracing infrastructure |
| Safety | Multi-tier content moderation with OpenAI Moderation API |
👉 Explore all models, datasets, and spaces at developer-lunark
-
wsbagnsv1/SkyReels-V2-I2V-14B-540P-GGUF
Image-to-Video • 16B • Updated • 378 • 16 -
wsbagnsv1/SkyReels-V2-T2V-14B-720P-GGUF
Image-to-Video • 14B • Updated • 116 • 8 -
wsbagnsv1/SkyReels-V2-I2V-14B-720P-GGUF
Image-to-Video • 16B • Updated • 249 • 8 -
wsbagnsv1/SkyReels-V2-T2V-14B-540P-GGUF
Text-to-Video • 14B • Updated • 166 • 3
-
wsbagnsv1/SkyReels-V2-I2V-14B-540P-GGUF
Image-to-Video • 16B • Updated • 378 • 16 -
wsbagnsv1/SkyReels-V2-T2V-14B-720P-GGUF
Image-to-Video • 14B • Updated • 116 • 8 -
wsbagnsv1/SkyReels-V2-I2V-14B-720P-GGUF
Image-to-Video • 16B • Updated • 249 • 8 -
wsbagnsv1/SkyReels-V2-T2V-14B-540P-GGUF
Text-to-Video • 14B • Updated • 166 • 3
datasets
0
None public yet