jetson-assistant-stage2-adapter

Stage 2 ("personal alignment") LoRA adapter from the llm_distillery project.

Trained on top of Stage 1's distilled knowledge (jetson-assistant-stage1-adapter merged into the base, then wrapped in a fresh LoRA — not stacked as a second adapter) on a small handwritten dataset (~60 examples) of personal-assistant behaviors: message drafting, calendar/scheduling, contact lookup, group-chat summarizing, and daily-workflow utilities.

Training: LoRA r=16/alpha=32 on all linear projections (q/k/v/o/gate/up/down proj), lr=5e-5, 2 epochs (~30 optimizer steps), max_length=512, FP16. Deliberately gentle settings — the dataset is tiny, and the goal is to layer on assistant behavior without disturbing Stage 1's reasoning/coding gains.

Note: this adapter's base_model_name_or_path metadata points at the plain Qwen2.5-0.5B base (a PEFT config quirk), but it was actually trained against Stage 1's merged weights. Apply it on top of Stage 1 merged in, or just use the final combined checkpoint at jetson-assistant-0.5b directly.