zephyr-dpo-qlora-gpt4-5e-6

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the generation/GPT4 dataset. It achieves the following results on the evaluation set:

• Loss: 0.9318
• Rewards/chosen: -6.0570
• Rewards/rejected: -7.0161
• Rewards/accuracies: 0.6430
• Rewards/margins: 0.9590
• Rewards/margins Max: 4.7709
• Rewards/margins Min: -3.0068
• Rewards/margins Std: 2.6570
• Logps/rejected: -960.1845
• Logps/chosen: -890.2942
• Logits/rejected: -1.4367
• Logits/chosen: -1.4956

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: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• total_eval_batch_size: 16
• 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	Rewards/margins Max	Rewards/margins Min	Rewards/margins Std	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.3843	0.28	100	0.6754	-0.8222	-0.9456	0.6190	0.1235	0.8275	-0.5849	0.4661	-353.1436	-366.8085	-2.4707	-2.5001
0.1345	0.56	200	0.9115	-6.6830	-7.3423	0.6230	0.6594	4.1300	-2.9206	2.3615	-992.8121	-952.8915	-1.3494	-1.4037
0.0467	0.85	300	0.9371	-6.0411	-6.9811	0.6450	0.9400	4.7437	-3.0368	2.6550	-956.6877	-888.7039	-1.4320	-1.4909

Framework versions

• PEFT 0.7.1
• Transformers 4.39.0.dev0
• Pytorch 2.1.2+cu121
• Datasets 2.14.6
• Tokenizers 0.15.2