--- license: other tags: - lora - mlx - customer-support - placeholder-safe - llama-3.2 base_model: meta-llama/Llama-3.2-3B-Instruct library_name: mlx datasets: - bitext/Bitext-customer-support-llm-chatbot-training-dataset --- # Customer Support LoRA Adapter (Llama-3.2-3B-Instruct · MLX) This repository provides **LoRA adapters** fine-tuned on the [Bitext Customer Support dataset](https://huggingface.co/datasets/bitext/Bitext-customer-support-llm-chatbot-training-dataset). The goal is **placeholder-safe** answers (e.g., `{{Order Number}}`, `{{Website URL}}` **must remain unchanged**). - **Base model**: [`meta-llama/Llama-3.2-3B-Instruct`](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) - **Runtime**: Apple Silicon **MLX** - **This repo contains**: `adapters.safetensors` (**LoRA only**; no base weights) > You must accept/download the base model under its own license (Meta Llama 3.2 Community License). > This repo **does not** include base weights. --- ## Quickstart (MLX CLI) Install once: ```bash pip install -U mlx-lm Generate with adapter: python -m mlx_lm.generate \ --model meta-llama/Llama-3.2-3B-Instruct \ --adapter-path batuhne/customer-support-llm \ --prompt "Hello, please cancel order {{Order Number}}" \ --max-tokens 128 --temp 0.2 Interactive chat (optional): python -m mlx_lm.chat \ --model meta-llama/Llama-3.2-3B-Instruct \ --adapter-path batuhne/customer-support-llm Recommended system prompt Keep placeholders exactly: You are a concise, helpful customer support agent. Preserve placeholders exactly as given (e.g., {{Order Number}}, {{Website URL}}). Notes Trained for placeholder integrity; do not modify {{…}} tokens in outputs. Use MLX quantized base weights if you need lower RAM/VRAM usage on Apple Silicon. Limitations Adapter quality is tied to the base model and dataset scope (customer support). This is not a general safety or policy layer; review outputs before production use. License Base model license: Meta Llama 3.2 Community License (must be accepted separately).