pythia-1b-kto-iter0

This model is a fine-tuned version of DatPySci/pythia-1b-sft-full on the DatPySci/iter0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.2591
• Rewards/real: 0.0604
• Rewards/generated: -1.0267
• Rewards/accuracies: 0.9460
• Rewards/margins: 1.0871
• Logps/generated: -570.8114
• Logps/real: -468.1696
• Logits/generated: 0.2253
• Logits/real: -0.2820

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-07
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/real	Rewards/generated	Rewards/accuracies	Rewards/margins	Logps/generated	Logps/real	Logits/generated	Logits/real
0.2932	0.38	300	0.2962	0.0718	-0.7855	0.9220	0.8572	-568.3989	-468.0556	0.2554	-0.2530
0.2689	0.77	600	0.2591	0.0604	-1.0267	0.9460	1.0871	-570.8114	-468.1696	0.2253	-0.2820

Framework versions

• Transformers 4.38.1
• Pytorch 2.2.1
• Datasets 2.17.1
• Tokenizers 0.15.2