mistral-dpo
This model is a fine-tuned version of TheBloke/OpenHermes-2-Mistral-7B-GPTQ on the None dataset. It achieves the following results on the evaluation set:
- Loss: 0.8911
- Rewards/chosen: 0.5387
- Rewards/rejected: 0.4878
- Rewards/accuracies: 0.5096
- Rewards/margins: 0.0509
- Logps/rejected: -174.3804
- Logps/chosen: -178.5185
- Logits/rejected: -2.5028
- Logits/chosen: -2.5350
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: 1
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 2
- training_steps: 250
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Rewards/chosen | Rewards/rejected | Rewards/accuracies | Rewards/margins | Logps/rejected | Logps/chosen | Logits/rejected | Logits/chosen |
---|---|---|---|---|---|---|---|---|---|---|---|
0.6703 | 0.0 | 10 | 0.6842 | -0.0001 | -0.0268 | 0.5865 | 0.0267 | -179.5257 | -183.9063 | -2.4290 | -2.4720 |
0.7119 | 0.0 | 20 | 0.6751 | 0.1584 | 0.0990 | 0.5769 | 0.0594 | -178.2678 | -182.3211 | -2.4542 | -2.4988 |
0.647 | 0.0 | 30 | 0.6702 | 0.3569 | 0.2540 | 0.5769 | 0.1029 | -176.7180 | -180.3367 | -2.4886 | -2.5306 |
0.6748 | 0.0 | 40 | 0.6712 | 0.3439 | 0.2229 | 0.5288 | 0.1210 | -177.0292 | -180.4664 | -2.5206 | -2.5581 |
0.6513 | 0.0 | 50 | 0.6707 | 0.4403 | 0.2838 | 0.5577 | 0.1565 | -176.4200 | -179.5021 | -2.5608 | -2.5853 |
0.6103 | 0.0 | 60 | 0.6695 | 0.6831 | 0.4769 | 0.5577 | 0.2063 | -174.4892 | -177.0740 | -2.5719 | -2.5933 |
1.0313 | 0.01 | 70 | 0.6724 | 0.7062 | 0.5084 | 0.5577 | 0.1978 | -174.1739 | -176.8436 | -2.5543 | -2.5843 |
0.6876 | 0.01 | 80 | 0.6804 | 0.6995 | 0.5144 | 0.5385 | 0.1850 | -174.1135 | -176.9104 | -2.5443 | -2.5829 |
0.9661 | 0.01 | 90 | 0.6828 | 0.7118 | 0.5376 | 0.5385 | 0.1742 | -173.8821 | -176.7873 | -2.5479 | -2.5846 |
0.7354 | 0.01 | 100 | 0.6757 | 0.6765 | 0.5039 | 0.5577 | 0.1726 | -174.2186 | -177.1401 | -2.5399 | -2.5758 |
1.0127 | 0.01 | 110 | 0.7129 | 0.6089 | 0.4855 | 0.5288 | 0.1234 | -174.4033 | -177.8165 | -2.5464 | -2.5760 |
1.0366 | 0.01 | 120 | 0.7440 | 0.6068 | 0.4946 | 0.5481 | 0.1122 | -174.3115 | -177.8369 | -2.5516 | -2.5804 |
1.2145 | 0.01 | 130 | 0.7564 | 0.6521 | 0.5396 | 0.5673 | 0.1125 | -173.8620 | -177.3846 | -2.5608 | -2.5878 |
0.8342 | 0.01 | 140 | 0.7649 | 0.6639 | 0.5519 | 0.5385 | 0.1119 | -173.7388 | -177.2668 | -2.5547 | -2.5828 |
0.7402 | 0.01 | 150 | 0.7991 | 0.5831 | 0.4883 | 0.5 | 0.0948 | -174.3747 | -178.0745 | -2.5498 | -2.5775 |
0.7162 | 0.01 | 160 | 0.8396 | 0.6134 | 0.5474 | 0.5096 | 0.0659 | -173.7835 | -177.7718 | -2.5445 | -2.5713 |
0.9396 | 0.01 | 170 | 0.8573 | 0.5700 | 0.5144 | 0.5288 | 0.0556 | -174.1144 | -178.2057 | -2.5326 | -2.5629 |
0.5958 | 0.01 | 180 | 0.8708 | 0.5526 | 0.5017 | 0.5288 | 0.0509 | -174.2406 | -178.3789 | -2.5227 | -2.5540 |
0.7588 | 0.02 | 190 | 0.8865 | 0.5428 | 0.4977 | 0.5288 | 0.0450 | -174.2806 | -178.4775 | -2.5207 | -2.5493 |
0.7811 | 0.02 | 200 | 0.8933 | 0.5797 | 0.5429 | 0.5192 | 0.0368 | -173.8286 | -178.1080 | -2.5171 | -2.5434 |
0.5735 | 0.02 | 210 | 0.8907 | 0.5577 | 0.5174 | 0.5288 | 0.0403 | -174.0838 | -178.3279 | -2.5069 | -2.5366 |
0.7709 | 0.02 | 220 | 0.8886 | 0.5602 | 0.5167 | 0.5192 | 0.0435 | -174.0907 | -178.3035 | -2.5041 | -2.5361 |
0.4914 | 0.02 | 230 | 0.8884 | 0.5237 | 0.4766 | 0.5192 | 0.0471 | -174.4924 | -178.6684 | -2.5050 | -2.5375 |
0.739 | 0.02 | 240 | 0.8910 | 0.5281 | 0.4796 | 0.5192 | 0.0485 | -174.4621 | -178.6240 | -2.5027 | -2.5351 |
0.5743 | 0.02 | 250 | 0.8911 | 0.5387 | 0.4878 | 0.5096 | 0.0509 | -174.3804 | -178.5185 | -2.5028 | -2.5350 |
Framework versions
- PEFT 0.7.1
- Transformers 4.36.0
- Pytorch 2.0.1+cu117
- Datasets 2.15.0
- Tokenizers 0.15.0
- Downloads last month
- 2
Model tree for aritrasen/mistral-dpo
Base model
mistralai/Mistral-7B-v0.1
Finetuned
teknium/OpenHermes-2-Mistral-7B
Quantized
TheBloke/OpenHermes-2-Mistral-7B-GPTQ