liberalis-cogitator

Summary

liberalis-cogitator is a conversational instruction-tuning dataset by Locutusque containing ~515k cleaned rows (586k in the uncleaned stage-1 split) sourced primarily from airoboros-3.2 and other datasets. The data spans system/human/gpt conversation turns covering STEM and programming challenges, roleplay transcripts, creative exchanges, synthetic patient–therapist dialogues, and open-ended reasoning prompts. It was used to train the liberalis-cogitator-llama-3.1-8b model, which is described as embracing the philosophy that thought should wander without leash or muzzle. The dataset is approximately 1GB in size.

Keywords

instruction-tuning conversational airoboros uncensored roleplay stem reasoning sft

Topics

NLP / Instruction Tuning

Research notes

• No explicit license or detailed description on the dataset card itself; the README contains only dataset_info metadata. Description is inferred from the associated model card (liberalis-cogitator-llama-3.1-8b) and data inspection. The 'source' field shows airoboros-3.2 as the primary source.