zephyr-7b-dpo-full-gpt_consistent-reward-scale-1-rpo-gamma-05

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.0087
• Rewards/chosen: -0.0012
• Rewards/rejected: -0.1544
• Rewards/accuracies: 0.7672
• Rewards/margins: 0.1532
• Logps/rejected: -261.9635
• Logps/chosen: -285.2108
• Logits/rejected: -2.3836
• Logits/chosen: -2.4800

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: 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
0.012	0.1147	50	0.0113	0.0778	-0.0076	0.7026	0.0854	-247.2809	-277.3098	-2.5021	-2.5787
0.011	0.2294	100	0.0100	0.0419	-0.0781	0.7069	0.1200	-254.3327	-280.9024	-2.3370	-2.4392
0.0104	0.3440	150	0.0098	-0.0076	-0.1403	0.7198	0.1326	-260.5493	-285.8531	-2.3956	-2.4944
0.0096	0.4587	200	0.0093	0.0289	-0.1110	0.7931	0.1399	-257.6194	-282.1954	-2.4209	-2.5140
0.0094	0.5734	250	0.0089	-0.0113	-0.1675	0.7802	0.1562	-263.2694	-286.2172	-2.3630	-2.4591
0.0096	0.6881	300	0.0088	-0.0133	-0.1638	0.7845	0.1505	-262.9025	-286.4156	-2.3821	-2.4772
0.0096	0.8028	350	0.0087	-0.0056	-0.1617	0.7802	0.1561	-262.6906	-285.6459	-2.3851	-2.4814
0.0093	0.9174	400	0.0087	-0.0012	-0.1544	0.7672	0.1532	-261.9635	-285.2108	-2.3836	-2.4800

Framework versions

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