dril-instruct

putting dril in a jar for future generations.

Model Details

Model Description

a LoRA finetune on top of vicuna-13b-cocktail, made to act like dril's tweets when given an instruction to "Create a joke about X", where X is anything.

Developed by: lun-4, dither
Model type: LoRA
Language(s) (NLP): English
License: WTFPL
Finetuned from model: vicuna-13b-cocktail

Model Sources [optional]

Repository: https://github.com/lun-4/dril-instruct
Blogpost: https://l4.pm/wiki/Personal%20Wiki/AI%20stuff/dril-instruct.html
Demo: Nope

Uses

by shitposters, for shitposting, this isn't useful for any other purpose

Out-of-Scope Use

literally anything other than shitposting

Bias, Risks, and Limitations

this model was finetuned on cocktail, which was made from vicuna but without the "ethical guardrails" AKA "As an AI language model, I can't" responses

How to Get Started with the Model

get text-generation-webui
get the base model, vicuna-13b-cocktail in the models folder
put this in the loras folder
there is a lot of hacking that i had to do to make loras work with GPTQ quantized models on my machine. those hacks are not portable
use the "Create a joke about X" model template.

Training Details

Training Data

around 3K dril tweets (that's what snscrape could get, even though there's 12K reported by twitter), and some 10 or so hand-made instructions to the dril tweets

Training Procedure

see blogpost

Preprocessing [optional]

see blogpost pls

Training Hyperparameters

Training regime: int8

Evaluation

see blogpost pls

Environmental Impact

fuck if i know

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

Hardware Type: NVidia A100 80GB
Hours used: 1, not including 4 hours of trying to bang a training script together
Cloud Provider: RunPod
Compute Region: EU-Norway
Carbon Emitted: 0.1 kg CO2 eq. (0.47 if you include the "banging rocks together" step)