zephyr-7b-dpo-full-ultrabin-3-avg-logprob

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.6123
• Rewards/chosen: -0.3746
• Rewards/rejected: -0.6400
• Rewards/accuracies: 0.7266
• Rewards/margins: 0.2654
• Logps/rejected: -65.1168
• Logps/chosen: -38.4143
• Logits/rejected: 6.8350
• Logits/chosen: 5.8881

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-06
• train_batch_size: 8
• eval_batch_size: 8
• seed: 55
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 128
• 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
0.6406	0.2092	100	0.6437	-0.2937	-0.4611	0.6523	0.1674	-47.2310	-30.3306	3.0709	1.9625
0.6294	0.4184	200	0.6292	-0.2954	-0.5087	0.6797	0.2133	-51.9860	-30.4985	5.9054	5.0104
0.6134	0.6276	300	0.6231	-0.2592	-0.5006	0.7266	0.2415	-51.1801	-26.8737	7.0553	5.9221
0.6092	0.8368	400	0.6123	-0.3746	-0.6400	0.7266	0.2654	-65.1168	-38.4143	6.8350	5.8881

Framework versions

• Transformers 4.44.0.dev0
• Pytorch 2.1.2
• Datasets 2.20.0
• Tokenizers 0.19.1