zephyr-qwen2-7b-dpo
This model is a fine-tuned version of tanliboy/zephyr-qwen2-7b-sft on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:
- Loss: 0.5492
- Rewards/chosen: -0.3377
- Rewards/rejected: -0.7347
- Rewards/accuracies: 0.7120
- Rewards/margins: 0.3970
- Logps/rejected: -339.1164
- Logps/chosen: -334.2959
- Logits/rejected: 0.0096
- Logits/chosen: 0.1430
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: 2
- eval_batch_size: 2
- seed: 42
- distributed_type: multi-GPU
- num_devices: 8
- gradient_accumulation_steps: 8
- total_train_batch_size: 128
- total_eval_batch_size: 16
- 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 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.6439 | 0.2094 | 100 | 0.6410 | -0.0729 | -0.1774 | 0.6280 | 0.1045 | -283.3870 | -307.8225 | -0.0798 | 0.0658 |
| 0.5947 | 0.4187 | 200 | 0.5743 | -0.3018 | -0.6030 | 0.6800 | 0.3013 | -325.9524 | -330.7078 | -0.0418 | 0.1044 |
| 0.5465 | 0.6281 | 300 | 0.5582 | -0.2623 | -0.6201 | 0.7080 | 0.3578 | -327.6540 | -326.7564 | 0.0043 | 0.1374 |
| 0.5513 | 0.8375 | 400 | 0.5497 | -0.3456 | -0.7435 | 0.7040 | 0.3979 | -340.0029 | -335.0901 | 0.0097 | 0.1434 |
Framework versions
- Transformers 4.40.2
- Pytorch 2.3.0+cu121
- Datasets 2.19.1
- Tokenizers 0.19.1