BraylonDash's picture
End of training
a46b542 verified
metadata
license: mit
library_name: peft
tags:
  - alignment-handbook
  - generated_from_trainer
  - trl
  - dpo
base_model: microsoft/phi-2
datasets:
  - HuggingFaceH4/ultrafeedback_binarized
model-index:
  - name: phi-2-gpo-renew2-b0.001-extra-v2-i1
    results: []

phi-2-gpo-renew2-b0.001-extra-v2-i1

This model is a fine-tuned version of DUAL-GPO/phi-2-gpo-renew2-b0.001-i0 on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

  • Loss: 0.0388
  • Rewards/chosen: 0.0266
  • Rewards/rejected: -0.0126
  • Rewards/accuracies: 0.6070
  • Rewards/margins: 0.0392
  • Logps/rejected: -379.8497
  • Logps/chosen: -369.7509
  • Logits/rejected: -0.9196
  • Logits/chosen: -0.9539

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
  • 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
0.098 0.06 100 0.0533 -0.0029 -0.0036 0.4980 0.0007 -370.8433 -399.2503 -0.7225 -0.8171
0.094 0.13 200 0.0491 -0.0390 -0.0525 0.5525 0.0135 -419.6949 -435.2693 -1.0754 -1.1388
0.0898 0.19 300 0.0452 -0.0184 -0.0403 0.5780 0.0218 -407.5088 -414.7480 -1.0291 -1.0858
0.0731 0.26 400 0.0430 -0.0069 -0.0331 0.5970 0.0262 -400.2979 -403.1916 -0.9864 -1.0412
0.0787 0.32 500 0.0422 -0.0122 -0.0473 0.6070 0.0351 -414.4887 -408.4566 -1.0587 -1.0975
0.0742 0.38 600 0.0406 0.0135 -0.0175 0.6085 0.0309 -384.7105 -382.8363 -0.9872 -1.0246
0.0635 0.45 700 0.0401 0.0166 -0.0188 0.6095 0.0354 -386.0258 -379.6696 -0.9903 -1.0225
0.0881 0.51 800 0.0395 0.0250 -0.0102 0.6085 0.0352 -377.4323 -371.2672 -0.9658 -0.9975
0.0753 0.58 900 0.0393 0.0304 -0.0046 0.5990 0.0350 -371.7872 -365.8699 -0.9026 -0.9456
0.0922 0.64 1000 0.0390 0.0286 -0.0075 0.5990 0.0361 -374.7669 -367.7319 -0.8801 -0.9184
0.0703 0.7 1100 0.0389 0.0227 -0.0161 0.6000 0.0387 -383.3026 -373.6226 -0.9300 -0.9602
0.0746 0.77 1200 0.0388 0.0226 -0.0179 0.6050 0.0405 -385.1601 -373.7153 -0.8944 -0.9306
0.0925 0.83 1300 0.0387 0.0263 -0.0131 0.6030 0.0393 -380.3072 -370.0340 -0.9171 -0.9494
0.0863 0.9 1400 0.0387 0.0269 -0.0123 0.6055 0.0392 -379.5608 -369.4450 -0.9121 -0.9447
0.0904 0.96 1500 0.0386 0.0268 -0.0124 0.6045 0.0392 -379.6000 -369.4944 -0.9203 -0.9536

Framework versions

  • PEFT 0.7.1
  • Transformers 4.36.2
  • Pytorch 2.1.2
  • Datasets 2.14.6
  • Tokenizers 0.15.2