File size: 4,310 Bytes
ce72571 8ac6170 ce72571 8ac6170 ce72571 9baa155 ce72571 6839fbc ce72571 8ac6170 938df25 ce72571 8ac6170 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 |
---
base_model: princeton-nlp/Llama-3-Base-8B-SFT
datasets:
- HuggingFaceH4/ultrafeedback_binarized
library_name: transformers
tags:
- alignment-handbook
- generated_from_trainer
pipeline_tag: text-generation
model-index:
- name: llama-3-8b-dpo-ultrafeedback-decrease_linear-1.0to0.95
results: []
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# llama-3-8b-dpo-ultrafeedback-decrease_linear-1.0to0.95
This model is released from the preprint: [DPO-Shift: Shifting the Distribution of Direct Preference Optimization](https://arxiv.org/abs/2502.07599). For more details, please refer to our [repository](https://github.com/Meaquadddd/DPO-Shift).
This model is a fine-tuned version of [princeton-nlp/Llama-3-Base-8B-SFT](https://huggingface.co/princeton-nlp/Llama-3-Base-8B-SFT) on the HuggingFaceH4/ultrafeedback_binarized dataset.
It achieves the following results on the evaluation set:
- Loss: 0.5619
- Rewards/chosen: -0.3784
- Rewards/rejected: -0.8957
- Dpo Lambda: 0.9528
- Rewards/accuracies: 0.7310
- Rewards/margins: 0.5173
- Logps/rejected: -360.6006
- Logps/chosen: -338.4835
- Logits/rejected: -1.0030
- Logits/chosen: -0.9672
## 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: 2
- gradient_accumulation_steps: 32
- total_train_batch_size: 128
- total_eval_batch_size: 4
- 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 | Dpo Lambda | Rewards/accuracies | Rewards/margins | Logps/rejected | Logps/chosen | Logits/rejected | Logits/chosen |
|:-------------:|:------:|:----:|:---------------:|:--------------:|:----------------:|:----------:|:------------------:|:---------------:|:--------------:|:------------:|:---------------:|:-------------:|
| 0.6826 | 0.1047 | 50 | 0.6803 | 0.0669 | 0.0399 | 0.9948 | 0.6690 | 0.0270 | -267.0431 | -293.9557 | -0.9094 | -0.8412 |
| 0.5951 | 0.2094 | 100 | 0.6223 | -0.0861 | -0.2745 | 0.9895 | 0.7130 | 0.1884 | -298.4850 | -309.2591 | -0.9195 | -0.8667 |
| 0.6296 | 0.3141 | 150 | 0.5972 | -0.2312 | -0.5289 | 0.9843 | 0.7100 | 0.2977 | -323.9177 | -323.7625 | -0.9008 | -0.8554 |
| 0.6219 | 0.4187 | 200 | 0.5784 | -0.4096 | -0.8051 | 0.9790 | 0.7310 | 0.3955 | -351.5381 | -341.6022 | -0.9313 | -0.8927 |
| 0.5738 | 0.5234 | 250 | 0.5685 | -0.4338 | -0.8864 | 0.9738 | 0.7260 | 0.4526 | -359.6707 | -344.0276 | -0.9691 | -0.9333 |
| 0.5598 | 0.6281 | 300 | 0.5695 | -0.4246 | -0.9086 | 0.9686 | 0.7220 | 0.4840 | -361.8922 | -343.1057 | -1.0002 | -0.9608 |
| 0.566 | 0.7328 | 350 | 0.5613 | -0.3470 | -0.8404 | 0.9633 | 0.7260 | 0.4934 | -355.0737 | -335.3493 | -0.9958 | -0.9592 |
| 0.5423 | 0.8375 | 400 | 0.5613 | -0.3837 | -0.8996 | 0.9581 | 0.7290 | 0.5159 | -360.9908 | -339.0213 | -1.0033 | -0.9665 |
| 0.5357 | 0.9422 | 450 | 0.5619 | -0.3784 | -0.8957 | 0.9528 | 0.7310 | 0.5173 | -360.6006 | -338.4835 | -1.0030 | -0.9672 |
### Framework versions
- Transformers 4.44.2
- Pytorch 2.4.0+cu121
- Datasets 2.21.0
- Tokenizers 0.19.1 |