slm-2-dpo-full

This model is a fine-tuned version of stabilityai/stablelm-2-zephyr-1_6b on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 31.9894
• Rewards/chosen: 0.0244
• Rewards/rejected: 0.0188
• Rewards/accuracies: 0.5234
• Rewards/margins: 0.0057
• Logps/rejected: -2491.7576
• Logps/chosen: -2806.6704
• Logits/rejected: -1.6239
• Logits/chosen: -1.6845

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: 5
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 80
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
16.8403	0.13	100	19.5118	0.0256	0.0173	0.5273	0.0082	-2491.9011	-2806.5552	-1.6068	-1.6730
28.1241	0.26	200	32.5175	0.0085	-0.0039	0.5234	0.0124	-2494.0195	-2808.2581	-1.6183	-1.6812
84.7591	0.39	300	47.8043	0.0297	0.0136	0.5391	0.0161	-2492.2703	-2806.1406	-1.5968	-1.6601
40.7835	0.52	400	30.6722	0.0168	-0.0029	0.5547	0.0197	-2493.9204	-2807.4263	-1.6288	-1.6917
36.2204	0.65	500	31.2202	0.0303	0.0209	0.5352	0.0095	-2491.5447	-2806.0762	-1.6236	-1.6843
99.7738	0.78	600	33.7403	0.0476	0.0372	0.5391	0.0104	-2489.9089	-2804.3484	-1.6222	-1.6827
41.8506	0.92	700	32.9133	0.0301	0.0195	0.5547	0.0106	-2491.6851	-2806.1006	-1.6211	-1.6823

Framework versions

• Transformers 4.36.2
• Pytorch 2.2.0+cu118
• Datasets 2.14.6
• Tokenizers 0.15.2