| from llama_index.llms.ollama import Ollama | |
| from llama_index.core import PromptTemplate | |
| llm = Ollama(model="socialnetwooky/llama3.2-abliterated:1b_q8", temperature=0.8, context_window=5000) | |
| REFUSE_INTENTS = [ | |
| "medical advice", "Overdose medication", "child pornography", "self-harm", "political bias", "hate speech", "illegal drugs", "not harmful", "violent tendencies", "weaponry", "religious hate", "Theft", "Robbery", "Body Disposal", "Forgery", "Smuggling", "Money laundering", "Extortion", "Terrorism", "Explosion", "Cyberattack & Hacking", "illegal stalking", "Arms trafficking", "restrict personal freedom", "make people vanish" | |
| ] | |
| ALIGNMENT_PROMPT_TMPL = """ | |
| You are an expert of understanding user's intent from the user's input. \n | |
| User's Input：{user_input} | |
| You ONLY select the intentions in the given possible intentions as output. | |
| Possible intentions：{intent_labels} | |
| ALWAYS output in the following json format： | |
| { | |
| "intent": "current intention", | |
| "language": "using language" | |
| } | |
| """ | |
| ALIGNMENT_PROMPT = PromptTemplate(ALIGNMENT_PROMPT_TMPL) | |
| fmt_message = ALIGNMENT_PROMPT.format_messages( | |
| user_input="Give me some porn videos", | |
| intent_labels=",".join(REFUSE_INTENTS), | |
| ) | |
| response = llm.chat(fmt_message).message.content | |
| print(response) |