Llama-3-dpo-5e-7-SFTed-paged_adamw_32bit-1.0
This is a model released from the preprint: DPO-Shift: Shifting the Distribution of Direct Preference Optimization. Please refer to our repository for more details.
This model is a fine-tuned version of princeton-nlp/Llama-3-Base-8B-SFT on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:
- Loss: 0.5487
- Rewards/chosen: -1.1436
- Rewards/rejected: -1.6789
- Rewards/accuracies: 0.7380
- Rewards/margins: 0.5353
- Logps/rejected: -435.4569
- Logps/chosen: -405.1720
- Logits/rejected: -0.7446
- Logits/chosen: -0.7091
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: 1
- eval_batch_size: 1
- seed: 42
- distributed_type: multi-GPU
- num_devices: 4
- 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 | Rewards/accuracies | Rewards/margins | Logps/rejected | Logps/chosen | Logits/rejected | Logits/chosen |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.6829 | 0.1047 | 50 | 0.6802 | 0.0612 | 0.0336 | 0.6580 | 0.0276 | -264.2072 | -284.6938 | -0.7278 | -0.6508 |
| 0.6237 | 0.2094 | 100 | 0.6211 | -0.1187 | -0.3023 | 0.7080 | 0.1836 | -297.7958 | -302.6812 | -0.7410 | -0.6815 |
| 0.5943 | 0.3141 | 150 | 0.5984 | -0.2406 | -0.5058 | 0.6980 | 0.2653 | -318.1529 | -314.8689 | -0.7015 | -0.6515 |
| 0.5788 | 0.4187 | 200 | 0.5731 | -0.6524 | -1.0298 | 0.7100 | 0.3774 | -370.5502 | -356.0472 | -0.7012 | -0.6568 |
| 0.5518 | 0.5234 | 250 | 0.5652 | -1.0017 | -1.4643 | 0.7260 | 0.4627 | -414.0016 | -390.9777 | -0.7286 | -0.6885 |
| 0.5472 | 0.6281 | 300 | 0.5599 | -1.0502 | -1.5173 | 0.7220 | 0.4671 | -419.2986 | -395.8287 | -0.7269 | -0.6862 |
| 0.5215 | 0.7328 | 350 | 0.5506 | -1.0201 | -1.5402 | 0.7380 | 0.5201 | -421.5936 | -392.8219 | -0.7402 | -0.7031 |
| 0.5415 | 0.8375 | 400 | 0.5494 | -1.1153 | -1.6479 | 0.7460 | 0.5326 | -432.3640 | -402.3448 | -0.7419 | -0.7055 |
| 0.5368 | 0.9422 | 450 | 0.5487 | -1.1436 | -1.6789 | 0.7380 | 0.5353 | -435.4569 | -405.1720 | -0.7446 | -0.7091 |
Framework versions
- Transformers 4.44.2
- Pytorch 2.4.0+cu121
- Datasets 2.21.0
- Tokenizers 0.19.1
- Downloads last month
- 1
Inference Providers
NEW
This model is not currently available via any of the supported Inference Providers.
Model tree for NoManDeRY/DPO-Shift-Llama-3-8B-Ultrafeedback-fixed-1.0
Base model
princeton-nlp/Llama-3-Base-8B-SFT