Ornith-1.0-35B-SharkQuant

Ornith-1.0-35B fine-tuned (LoRA, merged) as a SharkQuant desk agent: tool-calling (get_gex etc.), GEX/gamma-exposure interpretation, equity swing/portfolio reasoning, plus NVIDIA Nemotron agentic/coding/fundamentals data.

• SFT-LoRA on ~18.5k rows: NVIDIA Nemotron (tool-calling, SWE, finance) + SharkQuant domain data (real 2024 GEX feature->outcome, swing signals, discovery-pattern tool use).
• Loss 2.76 -> 0.45, token-acc 89.8%. LoRA on full-attention q/k/v/o layers (hybrid-attn model).
• Convention: positive net GEX = dealers long gamma (vol-suppressive, mean-reverting).

Inference: temperature 0.6-1.0, top_p 0.95. Emits <think> reasoning + <tool_call> XML.