metadata
base_model: alignment-handbook/zephyr-7b-sft-full
datasets:
- HuggingFaceH4/ultrafeedback_binarized
library_name: peft
license: apache-2.0
tags:
- alignment-handbook
- trl
- dpo
- generated_from_trainer
model-index:
- name: zephyr-7b-dpo-lora-r16-20k
results: []
zephyr-7b-dpo-lora-r16-20k
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: 0.5367
- Rewards/chosen: -0.7912
- Rewards/rejected: -1.4787
- Rewards/accuracies: 0.7103
- Rewards/margins: 0.6874
- Logps/rejected: -395.8989
- Logps/chosen: -362.3625
- Logits/rejected: -2.5102
- Logits/chosen: -2.5539
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-06
- train_batch_size: 4
- eval_batch_size: 8
- seed: 42
- distributed_type: multi-GPU
- gradient_accumulation_steps: 4
- total_train_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 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.6895 | 0.08 | 100 | 0.6896 | 0.0099 | 0.0028 | 0.6627 | 0.0072 | -247.7537 | -282.2447 | -2.8481 | -2.8901 |
| 0.653 | 0.16 | 200 | 0.6569 | -0.0133 | -0.0954 | 0.6865 | 0.0821 | -257.5692 | -284.5635 | -2.8339 | -2.8742 |
| 0.6385 | 0.24 | 300 | 0.6190 | -0.2742 | -0.4752 | 0.6905 | 0.2011 | -295.5536 | -310.6566 | -2.8031 | -2.8399 |
| 0.5689 | 0.32 | 400 | 0.6027 | -0.2972 | -0.5719 | 0.6944 | 0.2747 | -305.2159 | -312.9573 | -2.8083 | -2.8437 |
| 0.5689 | 0.4 | 500 | 0.5750 | -0.6614 | -1.0704 | 0.7242 | 0.4089 | -355.0662 | -349.3812 | -2.7152 | -2.7560 |
| 0.5884 | 0.48 | 600 | 0.5479 | -0.6965 | -1.2708 | 0.7123 | 0.5743 | -375.1053 | -352.8877 | -2.6322 | -2.6724 |
| 0.5366 | 0.56 | 700 | 0.5462 | -0.7254 | -1.3351 | 0.7123 | 0.6097 | -381.5439 | -355.7809 | -2.6144 | -2.6541 |
| 0.542 | 0.64 | 800 | 0.5451 | -0.6920 | -1.2686 | 0.7262 | 0.5766 | -374.8915 | -352.4363 | -2.5757 | -2.6163 |
| 0.5282 | 0.72 | 900 | 0.5412 | -0.7969 | -1.4275 | 0.7083 | 0.6306 | -390.7825 | -362.9279 | -2.5266 | -2.5716 |
| 0.5873 | 0.8 | 1000 | 0.5369 | -0.8233 | -1.5128 | 0.7083 | 0.6894 | -399.3072 | -365.5720 | -2.5254 | -2.5693 |
| 0.5152 | 0.88 | 1100 | 0.5384 | -0.7446 | -1.4196 | 0.7143 | 0.6749 | -389.9855 | -357.7025 | -2.5188 | -2.5620 |
| 0.5213 | 0.96 | 1200 | 0.5370 | -0.7888 | -1.4748 | 0.7063 | 0.6860 | -395.5133 | -362.1219 | -2.5135 | -2.5568 |
Framework versions
- PEFT 0.12.0
- Transformers 4.44.0
- Pytorch 2.4.0+cu121
- Datasets 2.20.0
- Tokenizers 0.19.1