zephyr-7b-dpo-full
This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the None dataset. It achieves the following results on the evaluation set:
- Loss: 0.7656
- Rewards/chosen: -3.8106
- Rewards/rejected: -6.8888
- Rewards/accuracies: 0.7405
- Rewards/margins: 3.0782
- Logps/rejected: -327.0232
- Logps/chosen: -316.3568
- Logits/rejected: -2.4373
- Logits/chosen: -2.3640
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: 4
- seed: 42
- distributed_type: multi-GPU
- num_devices: 4
- total_train_batch_size: 32
- total_eval_batch_size: 16
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.1
- 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 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.6612 | 0.05 | 100 | 0.6640 | 0.0464 | -0.0349 | 0.6565 | 0.0813 | -258.4842 | -277.7868 | -2.8352 | -2.7411 |
| 0.5924 | 0.1 | 200 | 0.6068 | 0.0778 | -0.2525 | 0.6927 | 0.3302 | -260.6598 | -277.4728 | -2.8153 | -2.7265 |
| 0.5488 | 0.15 | 300 | 0.5772 | 0.1688 | -0.4844 | 0.7385 | 0.6531 | -262.9787 | -276.5630 | -2.8364 | -2.7548 |
| 0.5144 | 0.2 | 400 | 0.5635 | 0.0609 | -0.7604 | 0.7347 | 0.8213 | -265.7392 | -277.6411 | -2.7890 | -2.7072 |
| 0.5399 | 0.25 | 500 | 0.5393 | 0.0316 | -0.9906 | 0.75 | 1.0221 | -268.0409 | -277.9347 | -2.8372 | -2.7565 |
| 0.5776 | 0.31 | 600 | 0.5706 | 0.0425 | -0.9799 | 0.7424 | 1.0224 | -267.9345 | -277.8257 | -2.8388 | -2.7569 |
| 0.5834 | 0.36 | 700 | 0.5596 | 0.0454 | -1.0216 | 0.7424 | 1.0670 | -268.3513 | -277.7964 | -2.7830 | -2.6941 |
| 0.5394 | 0.41 | 800 | 0.5358 | 0.0804 | -0.9536 | 0.7481 | 1.0341 | -267.6714 | -277.4460 | -2.6313 | -2.5480 |
| 0.5141 | 0.46 | 900 | 0.5412 | -0.2704 | -1.4309 | 0.7443 | 1.1605 | -272.4444 | -280.9546 | -2.6662 | -2.5832 |
| 0.51 | 0.51 | 1000 | 0.5350 | -0.2070 | -1.4043 | 0.7366 | 1.1973 | -272.1781 | -280.3206 | -2.7118 | -2.6217 |
| 0.5219 | 0.56 | 1100 | 0.5405 | -0.1673 | -1.3152 | 0.7290 | 1.1479 | -271.2871 | -279.9233 | -2.7451 | -2.6605 |
| 0.5391 | 0.61 | 1200 | 0.5320 | -0.2460 | -1.4452 | 0.7405 | 1.1992 | -272.5871 | -280.7106 | -2.7552 | -2.6692 |
| 0.536 | 0.66 | 1300 | 0.5502 | -0.1919 | -1.3564 | 0.7271 | 1.1645 | -271.6995 | -280.1697 | -2.7006 | -2.6126 |
| 0.6544 | 0.71 | 1400 | 0.5309 | -0.3757 | -1.6757 | 0.7080 | 1.3000 | -274.8926 | -282.0077 | -2.6970 | -2.6046 |
| 0.5697 | 0.76 | 1500 | 0.5662 | -0.2493 | -1.4791 | 0.7156 | 1.2297 | -272.9258 | -280.7440 | -2.7656 | -2.6730 |
| 0.5538 | 0.81 | 1600 | 0.5326 | -0.4658 | -1.6791 | 0.7214 | 1.2134 | -274.9264 | -282.9080 | -2.6934 | -2.5946 |
| 0.551 | 0.86 | 1700 | 0.5258 | -0.6217 | -1.8893 | 0.7137 | 1.2676 | -277.0278 | -284.4673 | -2.6535 | -2.5567 |
| 0.5708 | 0.92 | 1800 | 0.5639 | -0.5168 | -1.8962 | 0.7214 | 1.3794 | -277.0974 | -283.4186 | -2.6279 | -2.5564 |
| 0.5344 | 0.97 | 1900 | 0.5603 | -0.3788 | -1.8158 | 0.7271 | 1.4370 | -276.2931 | -282.0388 | -2.6680 | -2.5998 |
| 0.0925 | 1.02 | 2000 | 0.5587 | -0.4628 | -2.1277 | 0.7405 | 1.6648 | -279.4120 | -282.8788 | -2.6520 | -2.5825 |
| 0.112 | 1.07 | 2100 | 0.5731 | -0.6788 | -2.5908 | 0.7481 | 1.9120 | -284.0433 | -285.0383 | -2.5722 | -2.5094 |
| 0.0539 | 1.12 | 2200 | 0.5869 | -1.0820 | -2.9310 | 0.7366 | 1.8489 | -287.4448 | -289.0707 | -2.5937 | -2.5303 |
| 0.0811 | 1.17 | 2300 | 0.6306 | -0.8332 | -2.7204 | 0.7424 | 1.8872 | -285.3392 | -286.5822 | -2.5137 | -2.4560 |
| 0.0877 | 1.22 | 2400 | 0.5963 | -1.3075 | -3.3622 | 0.7481 | 2.0548 | -291.7576 | -291.3254 | -2.5925 | -2.5291 |
| 0.1114 | 1.27 | 2500 | 0.6126 | -1.3609 | -3.5524 | 0.7462 | 2.1915 | -293.6587 | -291.8594 | -2.4792 | -2.4142 |
| 0.0864 | 1.32 | 2600 | 0.6457 | -1.6093 | -3.7584 | 0.75 | 2.1491 | -295.7195 | -294.3440 | -2.5710 | -2.5058 |
| 0.0708 | 1.37 | 2700 | 0.6080 | -1.8094 | -3.7042 | 0.7462 | 1.8948 | -295.1769 | -296.3445 | -2.5394 | -2.4684 |
| 0.0794 | 1.42 | 2800 | 0.6010 | -1.7685 | -3.8603 | 0.7538 | 2.0918 | -296.7380 | -295.9354 | -2.5369 | -2.4663 |
| 0.1009 | 1.48 | 2900 | 0.6102 | -1.6050 | -3.5962 | 0.7347 | 1.9912 | -294.0973 | -294.3007 | -2.4834 | -2.4073 |
| 0.083 | 1.53 | 3000 | 0.6125 | -1.6395 | -3.6683 | 0.7424 | 2.0288 | -294.8184 | -294.6455 | -2.5306 | -2.4521 |
| 0.0871 | 1.58 | 3100 | 0.6392 | -1.7447 | -3.8250 | 0.75 | 2.0802 | -296.3850 | -295.6979 | -2.5032 | -2.4279 |
| 0.1168 | 1.63 | 3200 | 0.5973 | -1.6226 | -3.5602 | 0.7443 | 1.9376 | -293.7374 | -294.4764 | -2.5372 | -2.4606 |
| 0.0699 | 1.68 | 3300 | 0.5816 | -1.6383 | -3.5364 | 0.7424 | 1.8982 | -293.4994 | -294.6331 | -2.5287 | -2.4527 |
| 0.1082 | 1.73 | 3400 | 0.5895 | -1.8055 | -3.7976 | 0.7424 | 1.9920 | -296.1109 | -296.3059 | -2.5178 | -2.4442 |
| 0.09 | 1.78 | 3500 | 0.6231 | -1.8455 | -4.0234 | 0.75 | 2.1779 | -298.3694 | -296.7055 | -2.5261 | -2.4561 |
| 0.1238 | 1.83 | 3600 | 0.6047 | -1.6771 | -3.5997 | 0.7424 | 1.9226 | -294.1321 | -295.0213 | -2.6294 | -2.5512 |
| 0.0847 | 1.88 | 3700 | 0.5898 | -1.6725 | -3.5743 | 0.7347 | 1.9018 | -293.8779 | -294.9758 | -2.6224 | -2.5471 |
| 0.0908 | 1.93 | 3800 | 0.5817 | -1.6076 | -3.5381 | 0.7366 | 1.9304 | -293.5158 | -294.3269 | -2.5778 | -2.5047 |
| 0.0666 | 1.98 | 3900 | 0.6063 | -1.6950 | -3.7437 | 0.7309 | 2.0487 | -295.5718 | -295.2004 | -2.5784 | -2.5061 |
| 0.0173 | 2.03 | 4000 | 0.6213 | -2.1227 | -4.3451 | 0.7309 | 2.2224 | -301.5862 | -299.4778 | -2.6197 | -2.5495 |
| 0.0213 | 2.09 | 4100 | 0.6529 | -2.4461 | -4.9221 | 0.7366 | 2.4759 | -307.3557 | -302.7117 | -2.6029 | -2.5335 |
| 0.0149 | 2.14 | 4200 | 0.6934 | -3.0653 | -5.7847 | 0.7347 | 2.7194 | -315.9821 | -308.9039 | -2.5938 | -2.5272 |
| 0.0084 | 2.19 | 4300 | 0.7083 | -3.1845 | -6.0188 | 0.7405 | 2.8343 | -318.3230 | -310.0955 | -2.5088 | -2.4404 |
| 0.0059 | 2.24 | 4400 | 0.7193 | -3.3983 | -6.2807 | 0.7405 | 2.8824 | -320.9418 | -312.2334 | -2.5109 | -2.4479 |
| 0.0116 | 2.29 | 4500 | 0.7128 | -3.3425 | -6.1944 | 0.7462 | 2.8519 | -320.0795 | -311.6758 | -2.4787 | -2.4132 |
| 0.0077 | 2.34 | 4600 | 0.7219 | -3.2306 | -6.1475 | 0.7481 | 2.9169 | -319.6102 | -310.5562 | -2.4449 | -2.3779 |
| 0.0177 | 2.39 | 4700 | 0.7451 | -3.5469 | -6.5210 | 0.75 | 2.9742 | -323.3456 | -313.7194 | -2.3861 | -2.3174 |
| 0.0112 | 2.44 | 4800 | 0.7547 | -3.4801 | -6.4397 | 0.7424 | 2.9595 | -322.5316 | -313.0519 | -2.3939 | -2.3242 |
| 0.0071 | 2.49 | 4900 | 0.7691 | -3.8596 | -6.8490 | 0.7443 | 2.9895 | -326.6253 | -316.8460 | -2.3524 | -2.2834 |
| 0.0118 | 2.54 | 5000 | 0.7717 | -3.8862 | -6.8731 | 0.7462 | 2.9868 | -326.8659 | -317.1129 | -2.3347 | -2.2658 |
| 0.014 | 2.59 | 5100 | 0.7685 | -3.5970 | -6.5998 | 0.7481 | 3.0028 | -324.1335 | -314.2205 | -2.3512 | -2.2783 |
| 0.0208 | 2.64 | 5200 | 0.7741 | -3.9029 | -6.8895 | 0.7443 | 2.9866 | -327.0299 | -317.2794 | -2.3875 | -2.3143 |
| 0.0076 | 2.7 | 5300 | 0.7600 | -3.6159 | -6.5800 | 0.7424 | 2.9641 | -323.9353 | -314.4092 | -2.4331 | -2.3592 |
| 0.0146 | 2.75 | 5400 | 0.7768 | -3.7657 | -6.8555 | 0.7424 | 3.0898 | -326.6905 | -315.9074 | -2.4475 | -2.3751 |
| 0.0161 | 2.8 | 5500 | 0.7902 | -3.9170 | -7.0635 | 0.7481 | 3.1465 | -328.7701 | -317.4208 | -2.4332 | -2.3620 |
| 0.0056 | 2.85 | 5600 | 0.7827 | -3.9513 | -7.0687 | 0.7424 | 3.1174 | -328.8217 | -317.7632 | -2.4313 | -2.3599 |
| 0.0083 | 2.9 | 5700 | 0.7741 | -3.8805 | -6.9708 | 0.7443 | 3.0903 | -327.8432 | -317.0560 | -2.4324 | -2.3598 |
| 0.0243 | 2.95 | 5800 | 0.7657 | -3.8176 | -6.8913 | 0.7405 | 3.0737 | -327.0486 | -316.4268 | -2.4355 | -2.3620 |
Framework versions
- Transformers 4.35.0
- Pytorch 2.1.1+cu121
- Datasets 2.14.6
- Tokenizers 0.14.1
- Downloads last month
- 13
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.
Model tree for yimingzhang/zephyr-7b-dpo-full
Base model
mistralai/Mistral-7B-v0.1
Finetuned
alignment-handbook/zephyr-7b-sft-full