zephyr-7b-dpo-full-gpt-reward-scale-01

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: 0.5383
• Rewards/chosen: -1.3950
• Rewards/rejected: -2.4464
• Rewards/accuracies: 0.7241
• Rewards/margins: 1.0514
• Logps/rejected: -490.2826
• Logps/chosen: -423.5039
• Logits/rejected: 1.2749
• Logits/chosen: -0.2527

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
0.674	0.1147	50	0.6644	-0.0456	-0.1442	0.6724	0.0986	-260.0650	-288.5642	-2.5039	-2.6023
0.5874	0.2294	100	0.5920	-0.9820	-1.5650	0.6810	0.5830	-402.1482	-382.2076	0.3008	-0.2226
0.5612	0.3440	150	0.5695	-1.4677	-2.3665	0.6897	0.8989	-482.2998	-430.7732	2.3140	1.4310
0.5427	0.4587	200	0.5523	-1.3469	-2.2624	0.7241	0.9156	-471.8922	-418.6947	0.9223	-0.3630
0.5474	0.5734	250	0.5430	-1.0958	-2.0370	0.6897	0.9412	-449.3501	-393.5861	0.9071	-0.4403
0.5556	0.6881	300	0.5404	-1.3959	-2.3862	0.7198	0.9903	-484.2666	-423.5919	1.1950	-0.1993
0.5373	0.8028	350	0.5416	-1.5583	-2.5998	0.7284	1.0414	-505.6230	-439.8387	1.7159	0.2396
0.5405	0.9174	400	0.5383	-1.3950	-2.4464	0.7241	1.0514	-490.2826	-423.5039	1.2749	-0.2527

Framework versions

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