zephyr-dpop-qlora-uf-ours-5e-6-epoch1

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

• Loss: 1.6168
• Positive Losses: 8.8283
• Dpo Losses: 0.6446
• Rewards/chosen: 0.0003
• Rewards/rejected: -0.1296
• Rewards/accuracies: 0.6500
• Rewards/margins: 0.1299
• Rewards/margins Max: 0.6434
• Rewards/margins Min: -0.3494
• Rewards/margins Std: 0.3327
• Logps/rejected: -271.5416
• Logps/chosen: -284.5663
• Logits/rejected: -2.6717
• Logits/chosen: -2.7138

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	Positive Losses	Dpo Losses	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
0.6232	0.28	100	1.1413	4.2610	0.6656	0.0429	-0.0252	0.6230	0.0680	0.4086	-0.2281	0.2094	-261.0972	-280.3080	-2.6361	-2.6726
0.5625	0.56	200	1.7186	9.6677	0.6469	-0.0183	-0.1426	0.6420	0.1243	0.6362	-0.3433	0.3277	-272.8399	-286.4236	-2.6380	-2.6780
0.4748	0.85	300	1.6048	8.7062	0.6448	0.0014	-0.1274	0.6470	0.1288	0.6375	-0.3461	0.3295	-271.3224	-284.4528	-2.6733	-2.7153

Framework versions

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