Final deployable checkpoint from Phase 1 of the llm_distillery project — a personal AI assistant distilled and aligned entirely on a Jetson Orin Nano Super (8GB unified memory), start to finish.

Plain FP16 Qwen2ForCausalLM checkpoint — no PEFT/adapter machinery needed at inference, loads directly with AutoModelForCausalLM.from_pretrained("mohith-das/jetson-assistant-0.5b"). Built by merging two LoRA adapters sequentially into the Qwen2.5-0.5B base:

Stage 1 — cognitive distillation (jetson-assistant-stage1-adapter): reasoning, code, and conversational-structure imitation from teacher datasets (AM-DeepSeek-R1, OpenCodeInstruct, FineVision).
Stage 2 — personal alignment (jetson-assistant-stage2-adapter): trained on top of the Stage 1 merge, on a small handwritten dataset of assistant behaviors (message drafting, scheduling, contact lookup, summarizing).

Order matters: Stage 2 was trained on already-merged Stage 1 weights, so this checkpoint reproduces that exact order (base -> +stage1 -> +stage2), not an independent combination of both adapters.

This is also the base model for Phase 2 of the project — see jetson-hybrid-flat-cache-0.5b for a follow-on attention-architecture experiment (alternating sliding-window / gated-linear attention) built on top of this checkpoint.