Mistral-7B-Instruct-v0.2-multilingual-dpo-v1.0-v2

This model is a fine-tuned version of mistralai/Mistral-7B-Instruct-v0.2 on the nthakur/multilingual-ultrafeedback-binarized-dpo-v0.1, the nthakur/multilingual-distilabel-intel-orca-dpo-pairs-v0.1, the nthakur/multilingual-truthy-dpo-pairs-v0.1 and the nthakur/GSM8KInstruct-Parallel-instruct-dpo-v0.1 datasets. It achieves the following results on the evaluation set:

• Loss: 0.1324
• Rewards/chosen: -2.6738
• Rewards/rejected: -12.2394
• Rewards/accuracies: 0.9377
• Rewards/margins: 9.5656
• Logps/rejected: -1515.8665
• Logps/chosen: -607.0774
• Logits/rejected: 0.4952
• Logits/chosen: 0.3030

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: 0.0002
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 3
• gradient_accumulation_steps: 2
• total_train_batch_size: 24
• total_eval_batch_size: 12
• 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
0.2695	0.1361	500	0.2653	-0.4399	-4.5379	0.8680	4.0981	-745.7153	-383.6803	-1.3998	-1.5327
0.4349	0.2723	1000	0.3152	-2.6018	-7.1212	0.8515	4.5195	-1004.0471	-599.8698	4.1724	4.7868
0.531	0.4084	1500	0.4873	-2.4253	-8.0681	0.7855	5.6428	-1098.7278	-582.2241	-1.5195	-1.6538
0.1681	0.5446	2000	0.2003	-3.9555	-13.1169	0.9089	9.1613	-1603.6106	-735.2488	-0.1888	-0.3742
0.1778	0.6807	2500	0.2004	-3.4745	-11.9768	0.9242	8.5023	-1489.6012	-687.1464	-0.7118	-0.9608
0.1342	0.8169	3000	0.1452	-3.0928	-12.8477	0.9340	9.7549	-1576.6960	-648.9738	0.6727	0.5428
0.1252	0.9530	3500	0.1328	-2.7014	-12.3976	0.9383	9.6962	-1531.6849	-609.8344	0.5002	0.3026

Framework versions

• PEFT 0.7.1
• Transformers 4.41.2
• Pytorch 2.3.0+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1