metadata
license: mit
base_model: HuggingFaceH4/mistral-7b-sft-beta
tags:
- trl
- dpo
- generated_from_trainer
model-index:
- name: zephyr-7b-dpo-full
results: []
zephyr-7b-dpo-full
This model is a fine-tuned version of HuggingFaceH4/mistral-7b-sft-beta on the None dataset. It achieves the following results on the evaluation set:
- Loss: 0.0108
- Rewards/chosen: -5.9141
- Rewards/rejected: -7.7338
- Rewards/accuracies: 0.7266
- Rewards/margins: 1.8197
- Logps/rejected: -1030.7371
- Logps/chosen: -848.4521
- Logits/rejected: -1.6334
- Logits/chosen: -1.6493
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: 8
- gradient_accumulation_steps: 2
- total_train_batch_size: 128
- total_eval_batch_size: 64
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 5
Training results
| Training Loss | Epoch | Step | Validation Loss | Rewards/chosen | Rewards/rejected | Rewards/accuracies | Rewards/margins | Logps/rejected | Logps/chosen | Logits/rejected | Logits/chosen |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.2786 | 0.21 | 100 | 0.2781 | -0.0080 | -0.0710 | 0.6719 | 0.0631 | -264.4583 | -257.8367 | -2.7623 | -2.7774 |
| 0.1377 | 0.42 | 200 | 0.1450 | -0.5817 | -1.0385 | 0.6992 | 0.4567 | -361.2018 | -315.2145 | -2.7365 | -2.7512 |
| 0.1162 | 0.63 | 300 | 0.1186 | -1.0407 | -1.6725 | 0.7266 | 0.6318 | -424.5983 | -361.1053 | -2.4888 | -2.5058 |
| 0.1019 | 0.84 | 400 | 0.0997 | -1.6327 | -2.4828 | 0.7461 | 0.8501 | -505.6364 | -420.3094 | -2.2736 | -2.3013 |
| 0.0226 | 1.05 | 500 | 0.0406 | -2.9554 | -4.2565 | 0.7266 | 1.3012 | -683.0034 | -552.5746 | -2.1929 | -2.2303 |
| 0.0116 | 1.26 | 600 | 0.0298 | -3.0110 | -4.3717 | 0.7305 | 1.3607 | -694.5244 | -558.1376 | -2.1365 | -2.1643 |
| 0.0132 | 1.46 | 700 | 0.0320 | -2.8731 | -4.1217 | 0.7383 | 1.2486 | -669.5266 | -544.3542 | -2.1173 | -2.1453 |
| 0.0141 | 1.67 | 800 | 0.0285 | -2.8506 | -4.0446 | 0.7383 | 1.1939 | -661.8126 | -542.1040 | -2.0387 | -2.0557 |
| 0.008 | 1.88 | 900 | 0.0217 | -3.7087 | -4.9874 | 0.7148 | 1.2786 | -756.0888 | -627.9131 | -1.8927 | -1.9084 |
| 0.0015 | 2.09 | 1000 | 0.0135 | -4.8936 | -6.4137 | 0.7109 | 1.5202 | -898.7281 | -746.3977 | -1.7007 | -1.7103 |
| 0.0019 | 2.3 | 1100 | 0.0140 | -4.8675 | -6.4410 | 0.7188 | 1.5735 | -901.4539 | -743.7909 | -1.7341 | -1.7490 |
| 0.0014 | 2.51 | 1200 | 0.0128 | -5.1432 | -6.7584 | 0.7188 | 1.6152 | -933.1906 | -771.3603 | -1.7194 | -1.7313 |
| 0.0012 | 2.72 | 1300 | 0.0126 | -5.2094 | -6.8051 | 0.7227 | 1.5957 | -937.8638 | -777.9802 | -1.7283 | -1.7387 |
| 0.0012 | 2.93 | 1400 | 0.0126 | -5.3124 | -6.9529 | 0.7148 | 1.6405 | -952.6434 | -788.2790 | -1.7056 | -1.7185 |
| 0.0009 | 3.14 | 1500 | 0.0113 | -5.6394 | -7.3683 | 0.7188 | 1.7289 | -994.1813 | -820.9806 | -1.6707 | -1.6834 |
| 0.0007 | 3.35 | 1600 | 0.0115 | -5.6409 | -7.3656 | 0.7227 | 1.7247 | -993.9130 | -821.1270 | -1.6691 | -1.6823 |
| 0.0011 | 3.56 | 1700 | 0.0114 | -5.6893 | -7.4555 | 0.7227 | 1.7662 | -1002.9027 | -825.9682 | -1.6580 | -1.6727 |
| 0.0007 | 3.77 | 1800 | 0.0113 | -5.7534 | -7.5287 | 0.7227 | 1.7753 | -1010.2194 | -832.3766 | -1.6467 | -1.6620 |
| 0.0009 | 3.97 | 1900 | 0.0113 | -5.7308 | -7.5090 | 0.7227 | 1.7782 | -1008.2513 | -830.1171 | -1.6581 | -1.6731 |
| 0.0006 | 4.18 | 2000 | 0.0109 | -5.8887 | -7.6915 | 0.7266 | 1.8028 | -1026.5013 | -845.9089 | -1.6381 | -1.6538 |
| 0.0006 | 4.39 | 2100 | 0.0109 | -5.9096 | -7.7239 | 0.7266 | 1.8144 | -1029.7469 | -847.9958 | -1.6345 | -1.6501 |
| 0.0006 | 4.6 | 2200 | 0.0109 | -5.8953 | -7.7105 | 0.7266 | 1.8152 | -1028.4065 | -846.5691 | -1.6360 | -1.6516 |
| 0.0007 | 4.81 | 2300 | 0.0108 | -5.9141 | -7.7338 | 0.7266 | 1.8197 | -1030.7371 | -848.4521 | -1.6334 | -1.6493 |
Framework versions
- Transformers 4.35.2
- Pytorch 2.1.2+cu121
- Datasets 2.14.6
- Tokenizers 0.14.1