System-L Research

LLM pre-training, mid-training, and RL research, active since 2024, led by ShadOwO (previously known as ShadowLilac)

Focus: AI anime girl characters and streaming-optimized model development.

---

Research areas

• Continued pretraining around teaching LLMs VTuber/streamer/VRChat specific knowledge
• Reward models for character consistency and preference scoring
• Long-horizon RL around teaching LLMs to be first-person agents in videogames
• High quality low latency voice generation

---

Compute resources

The lab has access to the following compute resources. Additional compute is acquired via Lambda Cloud

• Nvidia GB10 (2x)
• i7-13700F, RTX 3090 (2x), 128gb RAM, 4TB
• 8TB Disk

Models, datasets, and research artifacts published here are part of ongoing work in character-driven LLM development.