tsavage68's picture
End of training
3d59382 verified
metadata
license: llama3
base_model: meta-llama/Meta-Llama-3-8B-Instruct
tags:
  - trl
  - dpo
  - generated_from_trainer
model-index:
  - name: MedQA_L3_250steps_1e6rate_05beat_CSFTDPO
    results: []

MedQA_L3_250steps_1e6rate_05beat_CSFTDPO

This model is a fine-tuned version of meta-llama/Meta-Llama-3-8B-Instruct on an unknown dataset. It achieves the following results on the evaluation set:

  • Loss: 0.5035
  • Rewards/chosen: -0.9245
  • Rewards/rejected: -2.5465
  • Rewards/accuracies: 0.7626
  • Rewards/margins: 1.6220
  • Logps/rejected: -26.4095
  • Logps/chosen: -20.0716
  • Logits/rejected: -0.9727
  • Logits/chosen: -0.9715

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: 1e-06
  • train_batch_size: 2
  • eval_batch_size: 1
  • seed: 42
  • gradient_accumulation_steps: 2
  • total_train_batch_size: 4
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: cosine
  • lr_scheduler_warmup_steps: 100
  • training_steps: 250

Training results

Training Loss Epoch Step Validation Loss Rewards/chosen Rewards/rejected Rewards/accuracies Rewards/margins Logps/rejected Logps/chosen Logits/rejected Logits/chosen
0.7075 0.0489 50 0.6367 0.2363 0.0705 0.6571 0.1658 -21.1755 -17.7501 -0.9379 -0.9373
0.6451 0.0977 100 0.6114 -0.8886 -1.7629 0.6923 0.8743 -24.8423 -19.9998 -0.9999 -0.9992
0.7372 0.1466 150 0.5770 -1.9159 -3.2984 0.7253 1.3825 -27.9133 -22.0544 -0.9880 -0.9871
0.4401 0.1954 200 0.5109 -0.9476 -2.5465 0.7516 1.5989 -26.4095 -20.1178 -0.9750 -0.9738
0.6774 0.2443 250 0.5035 -0.9245 -2.5465 0.7626 1.6220 -26.4095 -20.0716 -0.9727 -0.9715

Framework versions

  • Transformers 4.41.0
  • Pytorch 2.0.0+cu117
  • Datasets 2.19.1
  • Tokenizers 0.19.1