metadata

license: apache-2.0
base_model: alignment-handbook/zephyr-7b-sft-full
tags:
  - alignment-handbook
  - trl
  - dpo
  - 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: 0.4968
Rewards/chosen: -1.2029
Rewards/rejected: -2.2447
Rewards/accuracies: 0.7617
Rewards/margins: 1.0418
Logps/rejected: -487.1403
Logps/chosen: -382.8826
Logits/rejected: 1.6118
Logits/chosen: 0.6753

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
0.5632	0.2092	100	0.5684	-0.8769	-1.4450	0.7305	0.5681	-407.1669	-350.2800	-0.3489	-0.6128
0.5374	0.4184	200	0.5202	-0.7727	-1.5477	0.7852	0.7750	-417.4406	-339.8678	-0.1011	-0.6617
0.4826	0.6276	300	0.5018	-1.1013	-2.0815	0.7734	0.9802	-470.8159	-372.7218	1.3131	0.4518
0.495	0.8368	400	0.4969	-1.1626	-2.1853	0.7773	1.0227	-481.1959	-378.8522	1.4928	0.5873

Framework versions

Transformers 4.40.1
Pytorch 2.1.2+cu118
Datasets 2.19.0
Tokenizers 0.19.1