Model techniques map
Techniquesdata curationdata mixture & curriculum

specific method · filed under data curation

ProtocolQA-aligned RL training datasets

Tests analogous datasets during reinforcement learning to check for harmful uplift.

Also called Aligned RL datasets with ProtocolQA.

source
1
model
1
labs adopt it
0
strongest
evaluated

How sources treat it

One count per evidence span, weakest treatment to strongest.

evaluated 1

Documented in

Evidence

1 span quoted from the sources, strongest treatment first.

Tested analogous datasets during RL training to confirm no harmful uplift

evaluateddata curationin gpt-oss-120b and gpt-oss-20bOpenAI

Filed alongside

Other methods under data curation :: data mixture & curriculum.