Note The original Capybara dataset for SFT (ending with an assistant response)

Note The first iteration on Argilla's end to generate responses with `argilla/notus-7b-v1`, `teknium/OpenHermes-2.5-Mistral-7B`, and `mlabonne/NeuralBeagle14-7B`; then using GPT-4 as a judge via UltraFeedback using `distilabel`

Note A subset of `LDJnr/Capybara` dropping the duplicates, as apparently there are some duplicate entries within the Dove subset

Note A subset of `distilabel-internal-testing/Capybara-Deduped` removing the rows that have already been generated and judged in `argilla/distilabel-capybara-dpo-7k-binarized`

Note This subset contains the generations and preferences of the samples in `distilabel-internal-testing/Capybara-Deduped-Remaining`, and should be merged into `argilla/distilabel-capybara-dpo-7k-binarized`

Note The final dataset as an iteration on top of `LDJnr/Capybara` generating alternative completions with `argilla/notus-7b-v1`, `teknium/OpenHermes-2.5-Mistral-7B`, and `mlabonne/NeuralBeagle14-7B`; then using GPT-4 as a judge via UltraFeedback using `distilabel`