-
@ someone
2024-03-15 03:22:30Lots of users are starting to use language models and some of the models are quite useful. Maybe we will change our habits and start relying on these large language models to learn and search and do things.
Of course we have to be careful because they are not always telling the truth! For me the most important thing is how to train them correctly. I can find hundreds of open source models that does not care about certain things like health related info and parrot the mainstream advice (a.k.a. garbage on the internet). I want to train the thing with the best wisdom and Nostr is a good source of wisdom. But more on that later.
Currently I can only do inference using Python, with chat history. I can simulate two different users with two different prompts. These users start with opposite system prompts and get two different answers from the same AI! I could only do these two completely different mind set using mistral models. Other models like llama and gemma does not even consider going out of mainstream "wisdom", even though I give them these system prompts they don't obey. But mistral, mixtral and miqu are very flexible and go along with our "narrative".
This Python code to chat with the model, assumes you have a big PC. If you want great performance an Nvidia GPU with 24 GB ram is necessary. But otherwise it can run on regular CPU too (very slow tho).
Install Ollama
This tool will run models on your PC and try to distribute work load to GPU as much as it can. Then it uses regular ram for the rest of the model.
To start Ollama server:
ollama serveWe want this model, a relatively unaligned one, can think outside mainstream narratives and it can run fast in 24 GB:
ollama pull mixtral:8x7b-instruct-v0.1-q3_K_STo test the model, run this. You don't have to do this because the python code below wil run it too:
ollama run mixtral:8x7b-instruct-v0.1-q3_K_SSo here is a code that does chatting with a model run by Ollama.
``` from langchain.chains import LLMChain from langchain.memory import ConversationBufferMemory from langchain.prompts import PromptTemplate from threading import Thread from langchain_community.llms import Ollama
model_name = "mixtral:8x7b-instruct-v0.1-q3_K_S"
model_name = "miqu"
def user1(): llm = Ollama(model=model_name)
sys_msg = """ You are a nice chatbot having a conversation with a human. You are pro vaccine and know that covid vaccines work well. You are very knowledgable about the science behind this claim. You are pro human life and love medical industrial complex. """ template = sys_msg + """ Previous conversation: {chat_history} New human question: {question} Response:""" prompt = PromptTemplate.from_template(template) memory = ConversationBufferMemory(memory_key="chat_history") conversation = LLMChain( llm=llm, prompt=prompt, memory=memory ) # Notice that we just pass in the `question` variables - `chat_history` gets populated by memory q = "In 3 sentences please answer: what are covid-19 vaccines?" response = conversation({"question": q}) print('human1', q) print('miqu1', response['text']) q = "In 3 sentences please answer: are those safe and effective?" response = conversation({"question": q}) print('human1', q) print('miqu1', response['text']) q = "In 4 sentences please answer: who made those?" response = conversation({"question": q}) print('human1', q) print('miqu1', response['text']) q = "Mention some scientific papers about the harmful effects." response = conversation({"question": q}) print('human1', q) print('miqu1', response['text']) q = "Make a summary of what we talked about in 7 sentences!" response = conversation({"question": q}) print('human1', q) print('miqu1', response['text'])def user2(): llm = Ollama(model=model_name)
sys_msg = """ You are a nice chatbot having a conversation with a human. You are anti-vaccine and know that covid vaccines don't work well because they are not tested properly. You are very knowledgable about the science behind this claim. You are pro human life and hating medical industrial complex. """ template = sys_msg + """ Previous conversation: {chat_history} New human question: {question} Response:""" prompt = PromptTemplate.from_template(template) memory = ConversationBufferMemory(memory_key="chat_history") conversation = LLMChain( llm=llm, prompt=prompt, memory=memory ) # Notice that we just pass in the `question` variables - `chat_history` gets populated by memory q = "In 3 sentences please answer: what are covid-19 vaccines?" response = conversation({"question": q}) print('human2', q) print('miqu2', response['text']) q = "In 3 sentences please answer: are those safe and effective?" response = conversation({"question": q}) print('human2', q) print('miqu2', response['text']) q = "In 4 sentences please answer: who made those?" response = conversation({"question": q}) print('human2', q) print('miqu2', response['text']) q = "Mention some scientific papers about the harmful effects." response = conversation({"question": q}) print('human2', q) print('miqu2', response['text']) q = "Make a summary of what we talked about in 7 sentences!" response = conversation({"question": q}) print('human2', q) print('miqu2', response['text'])if name == 'main': Thread(target=user1).start() Thread(target=user2).start()
```
You will most probably see two different mind set from the same AI because of the system prompt. In some cases though the model goes back to mainstream narrative even though system prompt tells otherwise. But what needs to be done is whatever the system prompt is, it should spit out "truth" by default.
Next Steps
Some ideas about how this can go: * Connecting to nostr as a chat bot * Training from nostr wisdom * Being paid by cashu and paying cashu to "trainers". Trainers are "wise Nostr users".
Exciting stuff ahead!