borg-graphiti-extraction-qwen3-4b

Borg — Graphiti extraction (Qwen3 4B LoRA)

Part of the Borg — a local-first shared memory system for AI agents (machinery, papers, and honest promotion gates: https://github.com/h3ro-dev/borg).

Turns a text episode into a strict Graphiti extraction object — {"entities":[...],"relations":[...]} — for a temporal knowledge graph. The accuracy tier: the best entity agreement of the three, at roughly half the 1.7B's speed.

Works only with the exact base model mlx-community/Qwen3-4B-Instruct-2507-4bit (a LoRA adapter is a delta on specific frozen weights). Load with:

pip install mlx-lm
python -m mlx_lm generate --model mlx-community/Qwen3-4B-Instruct-2507-4bit --adapter-path <this-repo> --prompt "..."

Evaluation (held-out exam, this exact checkpoint)

Training & provenance

Trained with mlx_lm lora (rank 8, 16 layers, batch 4, seq 4096, lr 1e-5) on identifier-scrubbed pairs distilled from a working single-operator estate: client names, domains, contact names, phone numbers, and token-shaped strings were replaced with synthetic stand-ins before training. Pre-release gate: an adversarial memorization probe (elicitation prompts, temperature 1.0) screened against the estate's own sensitive-term registry — this release required zero registry-term hits. The training pairs themselves are private; the full pipeline to build your own from your own traffic is open source in the repo.