mistral-7b-orpo-alignment-handbook

This model is a fine-tuned version of mistralai/Mistral-7B-v0.1 on the argilla/dpo-mix-7k dataset. It achieves the following results on the evaluation set:

• Loss: 0.8562
• Rewards/chosen: -0.0394
• Rewards/rejected: -0.0485
• Rewards/accuracies: 0.6615
• Rewards/margins: 0.0091
• Logps/rejected: -0.9709
• Logps/chosen: -0.7882
• Logits/rejected: -2.9442
• Logits/chosen: -2.9335
• Nll Loss: 0.8317
• Log Odds Ratio: -0.6241
• Log Odds Chosen: 0.3600

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: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• total_eval_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• lr_scheduler_warmup_steps: 100
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen	Nll Loss	Log Odds Ratio	Log Odds Chosen
0.9081	0.95	100	0.8756	-0.0406	-0.0483	0.625	0.0077	-0.9657	-0.8116	-3.0351	-3.0266	0.8517	-0.6438	0.3078
0.8743	1.9	200	0.8544	-0.0391	-0.0474	0.6458	0.0083	-0.9474	-0.7823	-2.9519	-2.9423	0.8308	-0.6319	0.3327
0.7952	2.84	300	0.8562	-0.0394	-0.0485	0.6615	0.0091	-0.9709	-0.7880	-2.9507	-2.9399	0.8317	-0.6238	0.3606

Framework versions

• Transformers 4.39.0.dev0
• Pytorch 2.2.1+cu121
• Datasets 2.14.6
• Tokenizers 0.15.2