chat_400_STEPS_05beta_1e7rate_CDPOSFT

This model is a fine-tuned version of tsavage68/chat_600STEPS_1e8rate_SFT on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.6919
• Rewards/chosen: -0.0067
• Rewards/rejected: -0.0110
• Rewards/accuracies: 0.4308
• Rewards/margins: 0.0043
• Logps/rejected: -18.8240
• Logps/chosen: -16.7680
• Logits/rejected: -0.5983
• Logits/chosen: -0.5982

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-07
• train_batch_size: 4
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• training_steps: 400

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6929	0.0977	50	0.6947	-0.0000	0.0016	0.4066	-0.0016	-18.7989	-16.7547	-0.5985	-0.5983
0.694	0.1953	100	0.6903	0.0030	-0.0047	0.4527	0.0076	-18.8113	-16.7487	-0.5976	-0.5975
0.6941	0.2930	150	0.6914	-0.0038	-0.0090	0.4330	0.0052	-18.8200	-16.7622	-0.5986	-0.5985
0.6934	0.3906	200	0.6941	-0.0035	-0.0032	0.4044	-0.0003	-18.8084	-16.7616	-0.5978	-0.5977
0.6963	0.4883	250	0.6932	-0.0058	-0.0078	0.3890	0.0019	-18.8175	-16.7663	-0.5978	-0.5977
0.6898	0.5859	300	0.6900	-0.0039	-0.0122	0.4330	0.0084	-18.8264	-16.7624	-0.5980	-0.5979
0.6856	0.6836	350	0.6923	-0.0075	-0.0109	0.4571	0.0034	-18.8237	-16.7695	-0.5975	-0.5974
0.6978	0.7812	400	0.6919	-0.0067	-0.0110	0.4308	0.0043	-18.8240	-16.7680	-0.5983	-0.5982

Framework versions

• Transformers 4.40.1
• Pytorch 2.0.0+cu117
• Datasets 2.19.1
• Tokenizers 0.19.1