zephyr-7b-dpo-full-gpt_consistent-high-curriculum

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: 0.4943
• Rewards/chosen: -0.9906
• Rewards/rejected: -1.9162
• Rewards/accuracies: 0.7328
• Rewards/margins: 0.9257
• Logps/rejected: -438.1469
• Logps/chosen: -384.1485
• Logits/rejected: 1.5457
• Logits/chosen: 0.3365

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
0.6617	0.1147	50	0.6429	-0.0319	-0.1521	0.6983	0.1202	-261.7313	-288.2777	-2.4839	-2.5612
0.5671	0.2294	100	0.5741	-0.6263	-1.0517	0.6897	0.4254	-351.6960	-347.7197	-1.7192	-1.9024
0.5328	0.3440	150	0.5237	-0.5529	-1.2131	0.7155	0.6602	-367.8349	-340.3821	0.6797	-0.1069
0.5339	0.4587	200	0.5135	-0.8694	-1.6358	0.7284	0.7664	-410.0984	-372.0303	1.1593	0.1333
0.5206	0.5734	250	0.5051	-1.1478	-2.0209	0.7457	0.8731	-448.6093	-399.8708	2.1252	1.0597
0.5161	0.6881	300	0.4995	-1.0692	-1.9303	0.7414	0.8611	-439.5535	-392.0123	1.6077	0.4690
0.5113	0.8028	350	0.4953	-0.9905	-1.8979	0.7284	0.9073	-436.3081	-384.1431	1.4599	0.2827
0.5006	0.9174	400	0.4943	-0.9906	-1.9162	0.7328	0.9257	-438.1469	-384.1485	1.5457	0.3365

Framework versions

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