tags:
- json
pretty_name: Pipkin 10K DPO
size_categories:
- 1K<n<10K
Pipkin 10K DPO
This dataset is a DPO dataset sourced from a non-deduplicated version of the Pipkin dataset in which regenerations have been utilized to create DPO signals.
These are real data signals from real users.
It is not reccomended to utilize the prompt items of this dataset for SFT, as for the majaority of them, along with the responses are low quality. This works for DPO due to delta learning however does not transfer to SFT.
Format:
This dataset is in standard DPO format like so:
{
"prompt":[
{
"role":"user",
"content":"Where is Paris?"
}
],
"chosen":{
"role":"assistant",
"content": "Paris **is** in France."
},
"rejected":{
"role":"assistant",
"content":"It's possible that Paris is in France."
}
}
Standardization
The data is not directly standardized into a system->user->assistant->... pattern, it can have any role any where, including having an assistant role as the first non-system input.
To resolve this, it is reccomended to imply inject a user message containing a single period or another small piece of text in too match it.