metadata

license: apache-2.0
base_model: alignment-handbook/zephyr-7b-sft-full
tags:
  - alignment-handbook
  - generated_from_trainer
  - trl
  - dpo
  - generated_from_trainer
datasets:
  - HuggingFaceH4/ultrafeedback_binarized
model-index:
  - name: zephyr-7b-dpo-full
    results: []

zephyr-7b-dpo-full

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

Loss: 0.5046
Rewards/chosen: -1.1826
Rewards/rejected: -2.0581
Rewards/accuracies: 0.7246
Rewards/margins: 0.8756
Logps/rejected: -470.5493
Logps/chosen: -395.9858
Logits/rejected: 0.0457
Logits/chosen: -0.4473

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: 4
eval_batch_size: 4
seed: 42
distributed_type: multi-GPU
num_devices: 3
gradient_accumulation_steps: 10
total_train_batch_size: 120
total_eval_batch_size: 12
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
0.5764	0.2	100	0.5829	-0.3592	-0.7613	0.6931	0.4020	-340.8605	-313.6503	-2.4360	-2.4791
0.5169	0.39	200	0.5312	-0.8847	-1.6204	0.7066	0.7356	-426.7720	-366.2012	-0.8443	-1.2010
0.5133	0.59	300	0.5159	-1.1886	-1.9604	0.7246	0.7718	-460.7765	-396.5906	0.0460	-0.3853
0.4968	0.79	400	0.5058	-1.2445	-2.1063	0.7141	0.8618	-475.3639	-402.1766	0.2014	-0.2552
0.4833	0.98	500	0.5045	-1.1821	-2.0581	0.7260	0.8760	-470.5448	-395.9374	0.0436	-0.4496

Framework versions

Transformers 4.36.2
Pytorch 2.1.2
Datasets 2.14.6
Tokenizers 0.15.2