Ejafa's picture
Update README.md
d582bc6 verified
metadata
license: mit
base_model: microsoft/Phi-3-mini-128k-instruct
tags:
  - alignment-handbook
  - trl
  - dpo
  - generated_from_trainer
  - trl
  - dpo
  - generated_from_trainer
datasets:
  - princeton-nlp/llama3-ultrafeedback
model-index:
  - name: phi-3-mini-128k-instruct-simpo-lr-5e-07-gamma-1.5
    results: []

Description

This model was trained as part of the Reinforcement Learning - 24 project at Peking University, focusing on [simpo].

Authors

  • Ejafa Bassam
  • Yaroslav Ponomarenko

phi-3-mini-128k-instruct-simpo-lr-5e-07-gamma-1.5

This model is a fine-tuned version of microsoft/Phi-3-mini-128k-instruct on the princeton-nlp/llama3-ultrafeedback dataset. It achieves the following results on the evaluation set:

  • Loss: 1.6226
  • Rewards/chosen: -2.2430
  • Rewards/rejected: -2.6527
  • Rewards/accuracies: 0.625
  • Rewards/margins: 0.4097
  • Logps/rejected: -1.0611
  • Logps/chosen: -0.8972
  • Logits/rejected: 2.0148
  • Logits/chosen: 2.0096

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: 4
  • seed: 42
  • distributed_type: multi-GPU
  • num_devices: 8
  • gradient_accumulation_steps: 8
  • total_train_batch_size: 128
  • total_eval_batch_size: 32
  • 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
1.6417 0.8549 400 1.6236 -2.2390 -2.6457 0.6210 0.4067 -1.0583 -0.8956 2.0190 2.0146

Framework versions

  • Transformers 4.41.2
  • Pytorch 2.3.1+cu121
  • Datasets 2.20.0
  • Tokenizers 0.19.1