mzhaoshuai/Llama-3-8B-Instruct-refalign

Model

stars

commits

linked in READMEs

Oct 16, 2025

updated

conversational

endpoints_compatible

llama

safetensors

text-generation

text-generation-inference

transformers

Browse cluster: Large Language Model Fine-tuning & Quantization →

README

RefAlign: RL with Similarity-based Rewards

GitHub repository: https://github.com/mzhaoshuai/RefAlign

Paper: Learning from Reference Answers: Versatile Language Model Alignment without Binary Human Preference Data.

The training data is mzhaoshuai/Llama-3.3-70B-Inst-awq_ultrafeedback_1in3.

When conducting Reinforcement Learning with Similarity-based Rewards, the reward function is BERTScore.

Hyper-Parameters	Value
LR	2.5e-6
Batch Size	512
Epoch	1
Prompt Length	600
Generation Length	1200
Advantage CLIP	0.08
Sampled Generations (K)	2
BertScore Model	bart-large-mnli

Contributors

mzhaoshuai

6 commits

nielsr

1 commits

Similar repos

mzhaoshuai/Mistral-7B-Instruct-v0.2-refalign

Model

mzhaoshuai/Llama-3-8B-Instruct-ref-simpo

Model

mzhaoshuai/alpaca-7b-ref-bertscore

Model

mzhaoshuai/Mistral-7B-Instruct-v0.2-ref-simpo

Model

mzhaoshuai/alpaca-7b-ref-meteor

Model

mzhaoshuai/Llama-2-7b-hf-conf-refalign

Model

mzhaoshuai/llama3-ultrafeedback-bertscore-bart-large-mnli

Dataset

mzhaoshuai/Llama-3.3-70B-Inst-awq_ultrafeedback_1in3

Dataset

Linked in READMEs

mzhaoshuai/RefAlign

RL with Similarity-based Rewards for Versatile Language Model Alignment.

README badge

README

RefAlign: RL with Similarity-based Rewards

GitHub repository: https://github.com/mzhaoshuai/RefAlign

Paper: Learning from Reference Answers: Versatile Language Model Alignment without Binary Human Preference Data.

The training data is mzhaoshuai/Llama-3.3-70B-Inst-awq_ultrafeedback_1in3.

When conducting Reinforcement Learning with Similarity-based Rewards, the reward function is BERTScore.

Hyper-Parameters	Value
LR	2.5e-6
Batch Size	512
Epoch	1
Prompt Length	600
Generation Length	1200
Advantage CLIP	0.08
Sampled Generations (K)	2
BertScore Model	bart-large-mnli