Datasetpublic

davidheineman/text-ppl-dolci-response-pool

text-ppl-dolci-response-pool multi-model response pools split out of davidheineman/text-ppl, sampled from allenai/Dolci-DPO-Model-Response-Pool one config per (model, dataset), named dolci_response_pool_{model}_{dataset}, keeping the gemma / gpt / qwen / olmo model families: from datasets import load_dataset ds = load_dataset('davidheineman/text-ppl-dolci-response-pool', 'dolci_response_pool_olmo2_13b_DaringAnteater_prefs_olmo2_7b', split='test') the test split is the… See the full description on the dataset page: https://huggingface.co/datasets/davidheineman/text-ppl-dolci-response-pool.

sourceHugging Faceupdated 20d agoView on Hugging Face

0likes335downloads

Dataset Card

text-ppl-dolci-response-pool

multi-model response pools split out of davidheineman/text-ppl, sampled from allenai/Dolci-DPO-Model-Response-Pool

one config per (model, dataset), named dolci_response_pool_{model}_{dataset}, keeping the gemma / gpt / qwen / olmo model families:

python

from datasets import load_dataset
ds = load_dataset('davidheineman/text-ppl-dolci-response-pool', 'dolci_response_pool_olmo2_13b_DaringAnteater_prefs_olmo2_7b', split='test')

the test split is the benchmark; a config only has a validation split when its (model, dataset) pair had a second full 40M tokens to give

text sets should be formatted like {prompt}{context}{continuation} where PPL is computed over {continuation} only — here {prompt} is the user prompt, {context} is empty and {continuation} is the model response