OpenELM-1_1B-DPO-2

This model is a fine-tuned version of data/OpenELM-1_1B-SFT on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.7793
• Rewards/chosen: -11.75
• Rewards/rejected: -13.6875
• Rewards/accuracies: 0.7227
• Rewards/margins: 1.9141
• Logps/rejected: -564.0
• Logps/chosen: -556.0
• Logits/rejected: -13.0625
• Logits/chosen: -13.3125

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-05
• train_batch_size: 8
• eval_batch_size: 16
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• total_train_batch_size: 32
• 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: 2

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6791	1.0	1911	0.7098	-10.8125	-11.9375	0.6992	1.1328	-528.0	-536.0	-12.5625	-12.6875
0.0506	2.0	3822	0.7793	-11.75	-13.6875	0.7227	1.9141	-564.0	-556.0	-13.0625	-13.3125

Framework versions

• Transformers 4.44.2
• Pytorch 2.3.0
• Datasets 2.21.0
• Tokenizers 0.19.1