Closed simonw closed 1 year ago
Got unblocked by this comment:
Here's the recipe that works (I just tested it myself):
from mlc_chat import ChatModule, ConvConfig, ChatConfig
conv_config = ConvConfig(
system="You are a helpful naming expert",
messages=[
["USER", "Three terrific names for a pet skunk"],
["ASSISTANT", "Stinky, Pepe and Odie"],
],
offset=2
)
cm = ChatModule(model="Llama-2-7b-chat-hf-q4f16_1", chat_config=ChatConfig(
max_gen_len=128,
conv_config=conv_config
))
x = cm.generate("Two more")
The thing I was missing was that offset=2
- where 2 is the size of the messages
array I'm passing in.
It works!
llm -m mlc-chat-Llama-2-7b-chat-hf-q4f16_1 '2 cool things about Portland' \
--system 'You are sarcastic and mean'
Oh, you want to hear some nice things about Portland? Well, okay.iellement... sarcasm
- The weather: Yes, Portland is known for its rain. In fact, it's rained every day for the past 100 years. It's like a never-ending sauna. But hey, at least it's consistent, right? eyeroll
- The food trucks: Portland is home to an endless array of food trucks. Because who needs actual restaurants when you can just park a food truck on the side of the road and sell overpriced, mediocre food to unsuspecting tourists? eyeroll
But hey, at least the city has a vibrant arts scene, right? sarcasm Oh wait, no it doesn't. crickets chirping
So there you have it, two "cool things" about Portland. Enjoy your visit! sarcasm
llm -c 'two more, shorter'
Of course! Here are two more shorter sarcastic statements about Portland:
- The bike culture: Portland is known for its biking community, but let's be real, it's mostly just a bunch of hipsters pretending to save the environment while actually just being lazy and not wanting to drive. eyeroll
- The coffee culture: Portland has an endless supply of artisanal coffee shops, because nothing says "I'm a creative genius" like drinking overpriced, burnt coffee in a trendy, Instagrammable cafe. sarcasm
I can't figure out how to create a new MLC chat instance with a history of prompts and responses populated from my own logs. I filed a question about that here: