Find open source LLM to generate verbal cue

[StephanAkkerman]

Paper: "our text generator, GPT-3 [6] (text-davinci-003, temp=0.5), generates the verbal cue".

A few example sentences:

• Write a short, catchy sentence that connects flashy and bottle. Also, the sentence must start with 'Imagine'.
  • Imagine a flashy bottle that stands out from the rest!
• Write a short, catchy sentence that connects hairdresser and freezer. Also, the sentence must start with 'Imagine'.
  • Imagine your hairdresser inside a freezer

See huggingface LLMs, need the following criteria:

• Lightweight (max 7B, preferably less)
• English
• Ability to generate a useful verbal cue, use multiple tests
• https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard

[StephanAkkerman]

IFEval and BBH are most important for the models

[StephanAkkerman]

We should try out these options:

• https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct
• https://huggingface.co/microsoft/Phi-3-mini-4k-instruct / https://huggingface.co/microsoft/Phi-3.5-mini-instruct
• https://huggingface.co/jpacifico/Chocolatine-3B-Instruct-DPO-Revised

[StephanAkkerman]

http://qwenlm.github.io/blog/qwen2.5-llm/

[StephanAkkerman]

This looks quite done for now. Maybe we can change the system prompt etc. and later we will be able to change the model easily after #25