Self-Exploring Language Models: Active Preference Elicitation for Online Alignment.

DPO-Zephyr-7B

This model is a fine-tuned version of HuggingFaceH4/mistral-7b-sft-beta using synthetic data based on on the HuggingFaceH4/ultrafeedback_binarized dataset.

Model description

• Model type: A 7B parameter Zephyr-based Self-Exploring Language Models (SELM).
• License: MIT

Results

	AlpacaEval 2.0 (LC WR)	MT-Bench (Average)
SELM-Zephyr-7B-iter-3	24.00	7.48
SELM-Zephyr-7B-iter-2	23.40	7.72
SELM-Zephyr-7B-iter-1	20.28	7.42
DPO-Zephyr-7B	14.45	7.28

Training hyperparameters

The following hyperparameters were used during training:

• alpha: 0.001
• beta: 0.01
• train_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 4
• total_train_batch_size: 256
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• num_epochs: 1

Framework versions

• Transformers 4.40.2
• Pytorch 2.1.2+cu121
• Datasets 2.14.6
• Tokenizers 0.19.1