phi-2-gpo-ultrafeedback-lora
This model is a fine-tuned version of lole25/phi-2-sft-ultrachat-lora on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:
- Loss: 0.0004
- Rewards/chosen: -0.0084
- Rewards/rejected: -0.0177
- Rewards/accuracies: 0.6700
- Rewards/margins: 0.0093
- Logps/rejected: -233.2047
- Logps/chosen: -261.0818
- Logits/rejected: 0.8824
- Logits/chosen: 0.7796
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: 4
- seed: 42
- distributed_type: multi-GPU
- num_devices: 4
- gradient_accumulation_steps: 4
- total_train_batch_size: 64
- total_eval_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: 2
Training results
Training Loss | Epoch | Step | Logits/chosen | Logits/rejected | Logps/chosen | Logps/rejected | Validation Loss | Rewards/accuracies | Rewards/chosen | Rewards/margins | Rewards/rejected |
---|---|---|---|---|---|---|---|---|---|---|---|
0.0026 | 0.21 | 100 | 0.8151 | 0.9175 | -260.2373 | -231.4896 | 0.0025 | 0.5080 | 0.0001 | 0.0006 | -0.0005 |
0.0023 | 0.42 | 200 | 0.8092 | 0.9120 | -260.3932 | -232.1152 | 0.0023 | 0.6560 | -0.0015 | 0.0053 | -0.0068 |
0.0022 | 0.63 | 300 | 0.7992 | 0.9022 | -260.9179 | -232.8447 | 0.0022 | 0.6700 | -0.0067 | 0.0073 | -0.0141 |
0.0021 | 0.84 | 400 | 0.7884 | 0.8914 | -261.1620 | -233.2157 | 0.0022 | 0.6640 | -0.0092 | 0.0086 | -0.0178 |
0.0022 | 1.05 | 500 | 0.7821 | 0.8853 | -261.1852 | -233.3614 | 0.0021 | 0.7100 | -0.0094 | 0.0098 | -0.0193 |
0.002 | 1.26 | 600 | 0.7815 | 0.8840 | -261.1207 | -233.2843 | 0.0021 | 0.6940 | -0.0088 | 0.0097 | -0.0185 |
0.0021 | 1.47 | 700 | 0.7790 | 0.8816 | -261.0788 | -233.2560 | 0.0021 | 0.7000 | -0.0083 | 0.0099 | -0.0182 |
0.0021 | 1.67 | 800 | 0.7781 | 0.8811 | -261.0643 | -233.2740 | 0.0021 | 0.6940 | -0.0082 | 0.0102 | -0.0184 |
0.0021 | 1.88 | 900 | 0.7806 | 0.8833 | -261.0922 | -233.2118 | 0.0021 | 0.6900 | -0.0085 | 0.0093 | -0.0178 |
Framework versions
- PEFT 0.7.1
- Transformers 4.36.2
- Pytorch 2.1.2+cu118
- Datasets 2.14.6
- Tokenizers 0.15.2
- Downloads last month
- 2
Model tree for lole25/phi-2-gpo-ultrafeedback-lora
Base model
microsoft/phi-2
Adapter
this model