dpo_pythia1b_hh_rlhf.yml_local_29-04-24_13-31-33_xxxxx

This model is a fine-tuned version of sophiex/pythia-1b-sft_hh_rlhf on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.6581
• Rewards/chosen: -0.1633
• Rewards/rejected: -0.3103
• Rewards/accuracies: 0.5971
• Rewards/margins: 0.1470
• Logps/rejected: -160.0996
• Logps/chosen: -160.0996

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: 1e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• total_eval_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 150
• num_epochs: 1
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen
0.6931	0.0	1	0.6931	0.0	0.0	0.0	0.0	-158.4665	-158.4665
0.6701	0.2	503	0.6745	-0.0602	-0.1382	0.5734	0.0779	-159.0690	-159.0690
0.6624	0.4	1006	0.6670	-0.0864	-0.1939	0.5862	0.1075	-159.3303	-159.3303
0.6587	0.6	1509	0.6612	-0.1043	-0.2301	0.5891	0.1259	-159.5091	-159.5091
0.6511	0.8	2012	0.6581	-0.1633	-0.3103	0.5971	0.1470	-160.0996	-160.0996

Framework versions

• PEFT 0.10.0
• Transformers 4.38.1
• Pytorch 2.1.2+cu121
• Datasets 2.17.1
• Tokenizers 0.15.2