tsavage68's picture
End of training
c911926 verified
metadata
license: llama3
base_model: tsavage68/MedQA_L3_1000steps_1e6rate_SFT
tags:
  - trl
  - dpo
  - generated_from_trainer
model-index:
  - name: MedQA_L3_250steps_1e7rate_05beta_CSFTDPO
    results: []

MedQA_L3_250steps_1e7rate_05beta_CSFTDPO

This model is a fine-tuned version of tsavage68/MedQA_L3_1000steps_1e6rate_SFT on an unknown dataset. It achieves the following results on the evaluation set:

  • Loss: 0.6492
  • Rewards/chosen: 0.3403
  • Rewards/rejected: 0.2334
  • Rewards/accuracies: 0.6857
  • Rewards/margins: 0.1070
  • Logps/rejected: -33.3881
  • Logps/chosen: -30.6478
  • Logits/rejected: -0.7314
  • Logits/chosen: -0.7307

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-07
  • 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
0.6857 0.0489 50 0.6947 -0.0249 -0.0232 0.4879 -0.0018 -33.9011 -31.3784 -0.7318 -0.7312
0.6799 0.0977 100 0.6734 0.3881 0.3450 0.6681 0.0432 -33.1649 -30.5522 -0.7330 -0.7323
0.6286 0.1466 150 0.6528 0.4844 0.3866 0.6813 0.0978 -33.0816 -30.3598 -0.7312 -0.7306
0.6183 0.1954 200 0.6449 0.3270 0.2107 0.7143 0.1163 -33.4334 -30.6745 -0.7312 -0.7305
0.6593 0.2443 250 0.6492 0.3403 0.2334 0.6857 0.1070 -33.3881 -30.6478 -0.7314 -0.7307

Framework versions

  • Transformers 4.41.1
  • Pytorch 2.0.0+cu117
  • Datasets 2.19.1
  • Tokenizers 0.19.1