metadata
library_name: transformers
license: other
base_model: qnguyen3/vylinh-qwen-3b-merged
tags:
- llama-factory
- full
- generated_from_trainer
model-index:
- name: vylinh-dpo-v4
results: []
vylinh-dpo-v4
This model is a fine-tuned version of qnguyen3/vylinh-qwen-3b-merged on the dpo_mix_vi dataset. It achieves the following results on the evaluation set:
- Loss: 0.4053
- Rewards/chosen: 2.9905
- Rewards/rejected: -1.6223
- Rewards/accuracies: 0.7143
- Rewards/margins: 4.6128
- Logps/rejected: -596.4308
- Logps/chosen: -706.0082
- Logits/rejected: -2.3770
- Logits/chosen: -2.8608
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: 4
- gradient_accumulation_steps: 4
- total_train_batch_size: 32
- total_eval_batch_size: 8
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 4.0
Training results
| Training Loss | Epoch | Step | Validation Loss | Rewards/chosen | Rewards/rejected | Rewards/accuracies | Rewards/margins | Logps/rejected | Logps/chosen | Logits/rejected | Logits/chosen |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.3646 | 0.9995 | 1368 | 0.3423 | 3.0067 | 0.1768 | 0.7143 | 2.8299 | -578.4399 | -705.8457 | -2.3914 | -2.8822 |
| 0.1542 | 1.9996 | 2737 | 0.3495 | 2.9122 | -0.7397 | 0.75 | 3.6519 | -587.6050 | -706.7914 | -2.3921 | -2.8787 |
| 0.0501 | 2.9998 | 4106 | 0.3891 | 3.0809 | -1.3165 | 0.7143 | 4.3974 | -593.3735 | -705.1045 | -2.3870 | -2.8676 |
| 0.0124 | 3.9978 | 5472 | 0.4053 | 2.9905 | -1.6223 | 0.7143 | 4.6128 | -596.4308 | -706.0082 | -2.3770 | -2.8608 |
Framework versions
- Transformers 4.45.2
- Pytorch 2.5.0+cu121
- Datasets 2.21.0
- Tokenizers 0.20.1