metadata

license: mit
base_model: ZhangShenao/SELM-Phi-3-mini-4k-instruct-iter-2
tags:
  - alignment-handbook
  - dpo
  - trl
  - selm
datasets:
  - HuggingFaceH4/ultrafeedback_binarized
model-index:
  - name: SELM-Phi-3-mini-4k-instruct-iter-3
    results: []

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

SELM-Phi-3-mini-4k-instruct-iter-3

This model is a fine-tuned version of ZhangShenao/SELM-Phi-3-mini-4k-instruct-iter-2 using synthetic data based on on the HuggingFaceH4/ultrafeedback_binarized dataset.

Model description

Model type: A 3.8B parameter Phi3-instruct-based Self-Exploring Language Models (SELM).
License: MIT

Results

	AlpacaEval 2.0 (LC WR)	MT-Bench (Average)
SELM-Phi-3-mini-4k-instruct-iter-3	27.98	8.32
SELM-Phi-3-mini-4k-instruct-iter-2	26.79	8.44
SELM-Phi-3-mini-4k-instruct-iter-1	27.33	8.37
Phi-3-mini-4k-instruct	23.05	8.12

Our model also ranks highly on WildBench! 🔥

Training hyperparameters

The following hyperparameters were used during training:

alpha: 0.001
beta: 0.01
train_batch_size: 4
seed: 42
distributed_type: multi-GPU
num_devices: 8
gradient_accumulation_steps: 4
total_train_batch_size: 128
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