scpo-reward-model-australia / release-settings.json
minsik-oh's picture
Release reproduced country-specific SCPO DeBERTa checkpoint
26d35c8 verified
Raw History Blame Contribute Delete
329 Bytes
{
"country": "Australia",
"shuffle_seed": 2,
"learning_rate": 1e-06,
"tau": 0.5,
"beta": 1.0,
"batch_size": 8,
"max_steps": 1024,
"method": "filtered_weighted",
"model_name": "OpenAssistant/reward-model-deberta-v3-base",
"split_seed": 0,
"model_seed": 0,
"trainer_seed": 42,
"trainer_data_seed": null
}