metadata
license: apache-2.0
base_model: mistralai/Mistral-7B-v0.1
tags:
- alignment-handbook
- trl
- orpo
- generated_from_trainer
- trl
- orpo
- alignment-handbook
- generated_from_trainer
datasets:
- HuggingFaceH4/ultrafeedback_binarized
model-index:
- name: zephyr-7b-sft-full-orpo
results: []
zephyr-7b-sft-full-orpo
This model is a fine-tuned version of mistralai/Mistral-7B-v0.1 on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:
- Loss: 1.3771
- Rewards/chosen: -0.1391
- Rewards/rejected: -0.1930
- Rewards/accuracies: 0.6528
- Rewards/margins: 0.0539
- Logps/rejected: -3.8602
- Logps/chosen: -2.7813
- Logits/rejected: -2.8670
- Logits/chosen: -2.8498
- Nll Loss: 1.3532
- Log Odds Ratio: -1.0480
- Log Odds Chosen: 1.2201
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: 2e-05
- 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: inverse_sqrt
- 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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.5668 | 0.1049 | 100 | 0.5843 | -0.0456 | -0.0529 | 0.6151 | 0.0073 | -1.0580 | -0.9113 | -3.3148 | -3.3082 | 0.5516 | -0.6530 | 0.2184 |
| 0.5676 | 0.2098 | 200 | 0.5726 | -0.0441 | -0.0532 | 0.625 | 0.0092 | -1.0644 | -0.8811 | -3.0026 | -2.9992 | 0.5359 | -0.6474 | 0.2850 |
| 0.5819 | 0.3146 | 300 | 0.5552 | -0.0439 | -0.0531 | 0.6290 | 0.0092 | -1.0620 | -0.8770 | -3.1424 | -3.1391 | 0.5202 | -0.6464 | 0.2830 |
| 0.5738 | 0.4195 | 400 | 0.5411 | -0.0422 | -0.0517 | 0.6290 | 0.0096 | -1.0346 | -0.8434 | -3.1026 | -3.1020 | 0.5047 | -0.6522 | 0.2961 |
| 0.5478 | 0.5244 | 500 | 0.5319 | -0.0421 | -0.0525 | 0.6290 | 0.0105 | -1.0509 | -0.8415 | -3.0260 | -3.0286 | 0.4970 | -0.6382 | 0.3327 |
| 0.5146 | 0.6293 | 600 | 0.5240 | -0.0408 | -0.0508 | 0.6230 | 0.0100 | -1.0165 | -0.8165 | -3.1325 | -3.1275 | 0.4883 | -0.6418 | 0.3121 |
| 0.5298 | 0.7341 | 700 | 0.5188 | -0.0413 | -0.0541 | 0.6429 | 0.0128 | -1.0827 | -0.8267 | -3.0761 | -3.0755 | 0.4842 | -0.6219 | 0.3869 |
| 0.5181 | 0.8390 | 800 | 0.5141 | -0.0410 | -0.0524 | 0.6329 | 0.0114 | -1.0475 | -0.8198 | -3.1382 | -3.1394 | 0.4803 | -0.6322 | 0.3506 |
| 0.5239 | 0.9439 | 900 | 0.5086 | -0.0402 | -0.0506 | 0.6310 | 0.0104 | -1.0129 | -0.8045 | -3.1191 | -3.1171 | 0.4748 | -0.6328 | 0.3268 |
| 0.2888 | 1.0488 | 1000 | 0.5400 | -0.0436 | -0.0556 | 0.6429 | 0.0120 | -1.1128 | -0.8724 | -3.0171 | -3.0190 | 0.5058 | -0.6318 | 0.3794 |
| 0.29 | 1.1536 | 1100 | 0.5385 | -0.0437 | -0.0574 | 0.6468 | 0.0138 | -1.1487 | -0.8736 | -3.0027 | -3.0029 | 0.5042 | -0.6256 | 0.4247 |
| 0.2826 | 1.2585 | 1200 | 0.5428 | -0.0443 | -0.0581 | 0.6429 | 0.0139 | -1.1626 | -0.8854 | -2.9620 | -2.9583 | 0.5084 | -0.6254 | 0.4215 |
| 0.2796 | 1.3634 | 1300 | 0.5393 | -0.0441 | -0.0589 | 0.6468 | 0.0147 | -1.1771 | -0.8825 | -2.9256 | -2.9285 | 0.5060 | -0.6208 | 0.4508 |
| 0.2784 | 1.4683 | 1400 | 0.5365 | -0.0444 | -0.0589 | 0.6528 | 0.0145 | -1.1784 | -0.8885 | -2.9583 | -2.9594 | 0.5037 | -0.6236 | 0.4410 |
| 0.2873 | 1.5732 | 1500 | 0.5330 | -0.0436 | -0.0579 | 0.6448 | 0.0143 | -1.1584 | -0.8718 | -2.9664 | -2.9657 | 0.5004 | -0.6226 | 0.4364 |
| 0.276 | 1.6780 | 1600 | 0.5367 | -0.0442 | -0.0594 | 0.6409 | 0.0152 | -1.1879 | -0.8833 | -2.9358 | -2.9324 | 0.5041 | -0.6160 | 0.4570 |
| 0.2715 | 1.7829 | 1700 | 0.5349 | -0.0436 | -0.0580 | 0.6448 | 0.0145 | -1.1603 | -0.8710 | -3.0209 | -3.0194 | 0.5024 | -0.6272 | 0.4425 |
| 0.2717 | 1.8878 | 1800 | 0.5341 | -0.0450 | -0.0616 | 0.6548 | 0.0166 | -1.2325 | -0.8997 | -2.9579 | -2.9563 | 0.5023 | -0.6184 | 0.4824 |
| 0.2857 | 1.9927 | 1900 | 0.5408 | -0.0454 | -0.0620 | 0.6548 | 0.0166 | -1.2409 | -0.9088 | -3.0279 | -3.0350 | 0.5091 | -0.6193 | 0.4892 |
| 0.1137 | 2.0975 | 2000 | 0.6877 | -0.0620 | -0.0838 | 0.6706 | 0.0218 | -1.6761 | -1.2408 | -2.8815 | -2.8704 | 0.6539 | -0.6273 | 0.5767 |
| 0.1192 | 2.2024 | 2100 | 0.7577 | -0.0706 | -0.0981 | 0.6726 | 0.0275 | -1.9620 | -1.4122 | -2.8433 | -2.8372 | 0.7199 | -0.6210 | 0.6958 |
| 0.1178 | 2.3073 | 2200 | 1.1762 | -0.1205 | -0.1717 | 0.6528 | 0.0512 | -3.4342 | -2.4108 | -2.9107 | -2.8878 | 1.1197 | -0.7778 | 1.1628 |
| 0.1184 | 2.4122 | 2300 | 1.8520 | -0.1935 | -0.2541 | 0.6369 | 0.0606 | -5.0812 | -3.8696 | -2.9226 | -2.9102 | 1.7542 | -1.0562 | 1.3233 |
| 0.1172 | 2.5170 | 2400 | 1.0193 | -0.1001 | -0.1434 | 0.6409 | 0.0432 | -2.8671 | -2.0024 | -2.8710 | -2.8561 | 0.9736 | -0.8145 | 1.0075 |
| 0.1109 | 2.6219 | 2500 | 1.2050 | -0.1209 | -0.1677 | 0.6329 | 0.0468 | -3.3547 | -2.4183 | -2.8571 | -2.8457 | 1.1724 | -0.9768 | 1.0766 |
| 0.1238 | 2.7268 | 2600 | 2.6922 | -0.3036 | -0.3822 | 0.5873 | 0.0786 | -7.6444 | -6.0725 | -2.9967 | -2.9805 | 2.6498 | -1.6934 | 1.6674 |
| 0.1192 | 2.8317 | 2700 | 1.2391 | -0.1189 | -0.1634 | 0.625 | 0.0445 | -3.2671 | -2.3779 | -2.8836 | -2.8662 | 1.1910 | -0.9507 | 1.0201 |
| 0.1191 | 2.9365 | 2800 | 1.0214 | -0.0976 | -0.1394 | 0.6270 | 0.0418 | -2.7882 | -1.9523 | -2.8221 | -2.8059 | 0.9673 | -0.8558 | 0.9869 |
Framework versions
- Transformers 4.41.0.dev0
- Pytorch 2.3.0+cu121
- Datasets 2.19.1
- Tokenizers 0.19.1