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: 0.5168
Rewards/chosen: 0.1142
Rewards/rejected: -0.9329
Rewards/accuracies: 0.7695
Rewards/margins: 1.0471
Logps/rejected: -272.7659
Logps/chosen: -262.3270
Logits/rejected: -2.5421
Logits/chosen: -2.5184

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: 8
eval_batch_size: 8
seed: 42
distributed_type: multi-GPU
num_devices: 8
gradient_accumulation_steps: 2
total_train_batch_size: 128
total_eval_batch_size: 64
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
0.5155	0.21	100	0.5466	0.2422	-0.6585	0.7656	0.9007	-270.0222	-261.0468	-2.6271	-2.6117
0.4988	0.42	200	0.5294	0.0326	-0.9993	0.7617	1.0320	-273.4304	-263.1426	-2.5977	-2.5791
0.5354	0.63	300	0.5235	0.5272	-0.4961	0.7656	1.0232	-268.3980	-258.1974	-2.5449	-2.5231
0.5043	0.84	400	0.5181	0.1494	-0.9114	0.7734	1.0608	-272.5513	-261.9752	-2.5355	-2.5122

Framework versions

Transformers 4.36.2
Pytorch 2.1.2
Datasets 2.14.6
Tokenizers 0.15.2