This model is a fine-tuned version of pythia on the oa dataset. It achieves the following results on the evaluation set:
-
seed: 42
-
learning_rate: 5e-06
-
train_batch_size: 32
-
eval_batch_size: 8
-
optimizer: Adam with betas : {'lr': 5e-06, 'betas': [0.9, 0.999], 'eps': 1e-08, 'weight_decay': 0.0}
-
lr_scheduler_type: linear
-
training_steps: 5000
-
fp16
-
warmup_steps 5
-
Num examples = 53k
{
"epoch": 1.0,
"train_loss": 0.8031303182039198,
"train_runtime": 6338.6403,
"train_samples": 53455,
"train_samples_per_second": 8.433,
"train_steps_per_second": 0.264
}
