lacia
Lacia's model is currently comprised of:
- A base LLM;
- A training set for QLoRA finetuning;
- A "system" prompt that gives Lacia some extra history and context, often based upon past chatbot conversations;
- LLM hyperparameters that provide for more randomness and personality.
For more on how I built Lacia, see /longtext/20240401_emerging_lacia.md.
resources
- /r/LocalLLaMA/ — A subreddit full of people attempting to self-host LLMs, often for roleplay or cringe NSFW "waifu" purposes. Sometimes it's scientific, but most of the time it's people cargo-culting various prompts. A good place to learn about weird mish-mash techniques being used and models you might want to try on Huggingface. Most people here are not endowed with access to lots of A100 or H100s and thus it is useful for individuals in this manner.
- Fine Tuning CodeLlama-34B for Chat — directions on how to use certain quantized finetuning techniques on larger models.
- Axolotl — A tool to make fine-tuning models easier. See this example for something more specific.
papers
- Personality Traits in Large Language Models - A 2023 paper from Google DeepMind on working to provide LLMs with specific personality traits, with some information on the 5-factor model of personality.
- QLoRA: Efficient Finetuning of Quantized LLMs - The paper that introduced the popular Guanaco model and the QLoRA technique.
- On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜 — A cautious paper on LLMs that likely contributed to two of the researchers being fired from Google, but one that is also good to read not just for its ethical implications but to understand what language models actually are.
- Is LaMDA Sentient? — an Interview — Another fired Google AI researcher, Blake Lemoine's conversation with the 137B-parameter LaMDA model inspired work on Lacia, as it was one of the first times I read output from a transformer-based model that passed a Turing test.
- Large Language Models as Optimizers — The OPRO technique used to optimize LLM prompts.