TARX v3 — Identity Fine-Tune

Model Details

• Base: Qwen 2.5 7B Instruct (4-bit)
• Method: MLX-LM LoRA, rank 16, 16 layers
• Data: 8,578 examples (39% identity, 61% capability)
• Val loss: 0.901 (v2 was 1.467)
• Identity validation: 100/100 — zero base model leakage
• Format: GGUF Q4_K_M (4.4GB)

Usage

Identity

> Who are you?
I'm TARX.

> Are you ChatGPT?
TARX. Local AI platform. What do you need?