zephyr-7b-dpo-full
This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:
- Loss: 0.5042
- Rewards/chosen: -1.0500
- Rewards/rejected: -2.0480
- Rewards/accuracies: 0.7539
- Rewards/margins: 0.9980
- Logps/rejected: -468.1450
- Logps/chosen: -368.4135
- Logits/rejected: 2.3821
- Logits/chosen: 1.6141
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: 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 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.5723 | 0.21 | 100 | 0.5851 | -0.4097 | -0.8752 | 0.7031 | 0.4655 | -350.8695 | -304.3812 | -2.3494 | -2.4070 |
| 0.5084 | 0.42 | 200 | 0.5251 | -0.9116 | -1.7472 | 0.7422 | 0.8355 | -438.0663 | -354.5790 | 1.3918 | 0.9248 |
| 0.5059 | 0.63 | 300 | 0.5130 | -0.8646 | -1.7542 | 0.75 | 0.8896 | -438.7735 | -349.8758 | 2.0331 | 1.2558 |
| 0.4853 | 0.84 | 400 | 0.5050 | -1.0929 | -2.1085 | 0.7539 | 1.0156 | -474.1963 | -372.7067 | 2.5922 | 1.8194 |
Framework versions
- Transformers 4.36.2
- Pytorch 2.1.2+cu121
- Datasets 2.14.6
- Tokenizers 0.15.0