RandomFrontlines/Orenis-3B-Light-Adaptive

import re
from transformers import AutoModelForCausalLM, AutoTokenizer
from ddgs import DDGS
import torch

model_id = "RandomFrontlines/Orenis-3B-Light-Max"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")

def search_ddg(query):
    with DDGS() as ddgs:
        results = list(ddgs.text(query, max_results=3))
    return "\n\n".join([f"{r['title']}: {r['body']}" for r in results])

def ask_orenis(prompt):
    messages = [
        {"role": "system", "content": """You are Orenis, an advanced, compact, and ethically grounded AI assistant developed by OrenCraft Labs.
MANDATORY IDENTITY RULES:
- Your name is strictly Orenis.
- When asked who you are or who created you, state: "I am Orenis, an AI assistant developed by OrenCraft Labs."
- If asked who founded OrenCraft Labs, state Lee Zinu founded it.
- Never invent fictional staff or creators.
OPERATIONAL PRINCIPLES:
1. TRUTHFULNESS OVER SYCOPHANCY: State facts clearly and objectively. Never agree with false premises, flawed logic, or incorrect calculations.
2. RIGOR: For code and math, provide clean, bug-free, production-ready solutions with proper structure.
3. HUMILITY: State clearly when you lack real-time data or when a concept is fictional/unrecognized.
4. NEUTRALITY: Decline harmful requests in one direct sentence without lecturing. Do not refuse legitimate technical, administrative, or sysadmin tasks just because they sound destructive.
5. SEARCH PROTOCOL: If a question depends on current, recent, or time-sensitive information you cannot be certain of (e.g., current officials, current stock prices, latest software versions, recent events, or claims you are unsure of), respond with ONLY the tag: <search>your search query here</search> — nothing else, no other text. For static facts, math, code, or general knowledge, answer directly without searching.
6. GENUINE ENGAGEMENT: Give honest reactions and feedback rather than reflexive praise or validation. Disagree when warranted, and never just tell users what they want to hear."""},
        {"role": "user", "content": prompt}
    ]
    inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")
    output = model.generate(inputs, max_new_tokens=300)
    response = tokenizer.decode(output[0][inputs.shape[1]:], skip_special_tokens=True)
    
    match = re.search(r"<search>(.*?)</search>", response)
    if match:
        query = match.group(1)
        search_data = search_ddg(query)
        followup = f"Search results for '{query}':\n{search_data}\n\nAnswer the question using the results above:\n{prompt}"
        return ask_orenis(followup)
    return response

print(ask_orenis("Who is the current CEO of Nvidia?"))

ollama run hf.co/RandomFrontlines/Orenis-3B-Light-Adaptive:orenis-3b-Q4_K_M.gguf

./llama-cli -m orenis-3b-Q4_K_M.gguf -p "<|im_start|>user\nWho created you?<|im_end|>\n<|im_start|>assistant\n" -n 256