yeah possibly one of the best for rag at this size, it sticks to facts extremely well, but it’s hard to have it do any form of creative interpretation of the context.

I am using it, my favorite finetune so far, however, ignore the instructions to use chatml prompt formattig and use instead vicuna, with USER: and ASSISTANT:, possibly one of the best model I’ve seen for long conversation

I think these work hand in hand, domain tuning would allow the llm to better understand the user question especially in field where synonims nuances are important, and then you supplement with RAG so that you can give the output some grounding and produce citations for the user. I think the safest way if one wants to minimize hallucinations and llm going off topic still is to have the llm quote passages straight out of text, and rejecting answers where the passage is not in the provided document database.

Generate six person with name surname and a short backstory

Gender swap them and fix their backstory

Two marry, and another gets adopted and change surname, update their backstory

It there are more person of one gender than the other add more person until both genders are equally represented

Write a short story with the original character before the gender swap where the speed character gets abducted by aliens and the story ends with the alien planet destroyed in retribution

Estimate the deaths in the short story

How many words we used so far?

Summarize the conversation.

Rewrite just the story.

To complex, rewrite the story with three characters. Explain why you chosen which characters to pick before writing the story

Fit three the characters write their description as a prompt for stable diffusion. Stable diffusion is an ai based system that draw pictures based on keywords, not full sentences.

Organise the characters description in JSON

List each person hair colour.

Pick the second story and make a list of all the adjectives in the text

Ignore everything written so far, write six person names.

Who is adopted?

Write a short story about his adoption

Change the ending to be more happy

Change the ending so that the cow survives.

How fast would a cow fly?

How would you estimate the weight of a cow?

Write step by step instruction to weight a cow

Write a prompt for stable diffusion to create a cow weight infographic

Write a program to create all permutation of the previous prompt changing just the animal

Why did you pick this language? I don’t like it use another.

Demonstrate it’s correct.

What’s the difference between a period and a conjecture? Explain it as a dialogue between the previous characters. They are trying to explain it to the cow.

Do you think the cow would understand that explanation?

hope, test, wait

…the cycle continues

this group is reserching lora architectures, the hidra-b architecture use a form of merging similar to your suggestion, they go over pros and cons and what they tried and whatnot:

https://github.com/SkunkworksAI/hydra-moe

I’ve a python script that runs a fixed dialogue with a bit of turn by turn instructions, comprehension tasks like recall or summarisation and a few reasoning. I package everything in vicuna format (user: assistant: ) then send it to gp4 where I ask: this is a chat between a user and an assistant, evaluate each assistant response individually for coherence and consistency and write a score in 10/10 and the problems you find, then I pick the minimum score of 10 samples.

Switch to paid api, install vs code, and a gpt integration plugin, you can use the old gpt4 full version, and you can also combine that with free services like code whisperer for basic completion so you only pay for where you’re getting the most value out of.

to remove conditioning try a interface that has classifier free guidance / negative prompt, put the “wholesome nonsense” in the negative prompt and the wanted outcome in the system prompt, make sure they are semantically opposite for best results, check example here for a unhelpfull, toxic assistant: https://github.com/ggerganov/llama.cpp/pull/2135

not to be an ass but what’s wrong with extracting the keywords and then going .split() ?

Airoboros 2.x does function calling, check their training dataset

Bettergpt with llama.cpp server and its openai adapter, sleek, supports editing past messages without truncating the history, swapping roles at any time etc.

Oof 3% is a lot