zephyr-7b-dpo-full-magpi-high-bleu-3-epochs

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: 0.0050
• Rewards/chosen: -1.4582
• Rewards/rejected: -44.8746
• Rewards/accuracies: 0.9960
• Rewards/margins: 43.4164
• Logps/rejected: -5128.2480
• Logps/chosen: -512.8050
• Logits/rejected: -3.4441
• Logits/chosen: -3.5504

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: 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: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.0025	0.4739	50	0.0067	-1.4454	-37.1575	0.9940	35.7120	-4356.5356	-511.5262	-3.0434	-3.2142
0.0036	0.9479	100	0.0053	-2.0303	-41.2450	0.9940	39.2146	-4765.2842	-570.0164	-3.2429	-3.4104
0.0001	1.4218	150	0.0070	-1.9459	-45.2030	0.9940	43.2570	-5161.0879	-561.5757	-3.5068	-3.5867
0.0	1.8957	200	0.0047	-1.4539	-44.2686	0.9960	42.8147	-5067.6450	-512.3704	-3.4229	-3.5020
0.0	2.3697	250	0.0050	-1.4525	-44.7537	0.9960	43.3012	-5116.1577	-512.2269	-3.4445	-3.5510
0.0	2.8436	300	0.0050	-1.4582	-44.8746	0.9960	43.4164	-5128.2480	-512.8050	-3.4441	-3.5504

Framework versions

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