meow_d LLM, v1

link to some AI slop it generated.

dataset is acquired from my posts.json in my cohost-dl downloads, only because i already had it downloaded. it’s probably better to only use the official data export, which only contains your own posts for legal reasons.

data preprocessing is done with python:

• get rid of shares, which can be identifyed as their ["astmap"]["spans"] are empty
• get rid of all the data except for your title and post itself ("headline" and "plainTextBody")

then format the dataset. can be done in different ways:

• human: “write a cohost post”, assistant: title + post
• no prompt template, just title + post

choosing the model:

• “open source” models
  • when it comes to machine learning, the main bottleneck is vram, and training uses more vram than running.
  • train using unsloth, axoltol, etc
  • if you have a beefy computer or is using a relatively small model, you can just train on there. if you don’t, use Google Colab or any other service out there.
• proprietary models
  • you’ll only be able to use their proprietary services to train
  • example: Google AI studio

(written months ago i forgot about this so might as well put it out there. im too lazy to make it good)