LoRA v2: When Your Repo Expert Starts Compounding

LoRA v1 let you train once from chat history — import a Hugging Face adapter, install pack presets, or export channel/collab/repo transcripts to Unsloth and compose an Ollama tag. It worked. It also felt like a sidecar: one-shot training, manual base-tag confusion, collab-only routing, no rollback, no team path.

LoRA v2 closes the loop. Personal learning v2 handles prompt-time memory; LoRA v2 handles weight-time compounding — same Specialist tuning pack, same gates, same local Ollama stack.

v1 vs v2

Area	v1	v2
Training	Full export every time	Incremental refresh + curated rows
Lifecycle	Train when 10+ turns	Ready / refresh badges, version rollback
Models	User picks Llama base	Transparent dual-tag profile (Qwen chat, Llama compose)
Routing	Keyword rules, collab tasks	Unified classifier, chat + collab + impl
Pipeline	One in-memory job	Queue, MLX on Apple Silicon, post-train eval
Sharing	Manual HF upload	Publish + MCP manifest + import suggestions

Five pillars

1. Compound learning loop — Adapter registry at ~/.neural-junkie/lora-adapters.json. Refresh from delta rows instead of restarting. Roll back to a prior version with one API call.

2. Transparent dual-tag profiles — Agents store model_profile: inference on Qwen, compose on Llama, one composed tag in the UI. Tool loops still run on Qwen when the composed tag lacks native tools.

3. Smarter routing — One SelectLoRATag path for chat, collaboration, and planning. LLM classifier can emit lora_tag. Repo-aware preference when nj-repo-* is installed.

4. Training you can trust — FIFO queue with persisted jobs. make deps-lora-mlx on Apple Silicon. Post-train eval before assign. Publish adapters to Hugging Face from the hub.

5. MCP + team — Repo MCP exports include LoRA metadata. Import suggests train or compose. Teammates pull weights from HF while learnings travel as JSON bundles.

Hero workflow

v1: create-repo-agent → 10+ Q&A → Train LoRA → assign tag
v2: same expert compounds → learnings inject daily → refresh when delta ready
     → eval passes → profile updates → teammate imports MCP + HF adapter

What didn't change

Specialist tuning pack gates. User confirmation for training and learnings. Ollama-local compose. Prompt personas, context stack, collaboration gates unchanged.

Try it

make pull-models
make deps-lora
make deps-lora-mlx   # Apple Silicon

Enable Specialist tuning → train or refresh from agent info → Model library.

Docs: LORA_V2.md

Download: https://camronwood.github.io/neural-junkie/download.html

Neural Junkie is a personal open-source project. Feedback welcome on GitHub.