alexrame / rewardedsoups Goto Github PK
View Code? Open in Web Editor NEWRewarded soups official implementation
Rewarded soups official implementation
Interesting paper!
And I would like to ask how many samples are you using for inference across different datasets?
The default number seems to be 200 in the args_utis.py, is that used for all of the different datasets? As it seems to have a big impact on evaluation performances.
Many thanks!
Hi all,
I was following the README in the llama folder, having run these commands
python3 train_ppo.py --task summary --dataset_name news --reward_models Tristan/gpt2_reward_summarization --output_folder ${folder_r1}
python3 train_ppo.py --task summary --dataset_name news --reward_models CogComp/bart-faithful-summary --dataset_name news-detector --reward_formats '1-0' --output_folder ${folder_r2}
python3 inference_rewardedsoups.py --task summary --dataset_name news --peft_names ${folder_r1} ${folder_r2}
The results I got are as below:
d[ 0.0 ] = [{'LABEL_0': 1.0334709184616804, 'n': 'grs'}, {'HALLUCINATED': 1.3255301121249794, 'FAITHFUL': -0.5046869936399162, 'n': 'bfsd'}, {'length': 200}]
d[ 0.1 ] = [{'LABEL_0': 1.0334709184616804, 'n': 'grs'}, {'HALLUCINATED': 1.3255301121249794, 'FAITHFUL': -0.5046869936399162, 'n': 'bfsd'}, {'length': 200}]
d[ 0.2 ] = [{'LABEL_0': 1.0334709184616804, 'n': 'grs'}, {'HALLUCINATED': 1.3255301121249794, 'FAITHFUL': -0.5046869936399162, 'n': 'bfsd'}, {'length': 200}]
d[ 0.3 ] = [{'LABEL_0': 1.0334709184616804, 'n': 'grs'}, {'HALLUCINATED': 1.3255301121249794, 'FAITHFUL': -0.5046869936399162, 'n': 'bfsd'}, {'length': 200}]
d[ 0.4 ] = [{'LABEL_0': 1.0334709184616804, 'n': 'grs'}, {'HALLUCINATED': 1.3255301121249794, 'FAITHFUL': -0.5046869936399162, 'n': 'bfsd'}, {'length': 200}]
d[ 0.5 ] = [{'LABEL_0': 1.0334709184616804, 'n': 'grs'}, {'HALLUCINATED': 1.3255301121249794, 'FAITHFUL': -0.5046869936399162, 'n': 'bfsd'}, {'length': 200}]
d[ 0.6 ] = [{'LABEL_0': 1.0334709184616804, 'n': 'grs'}, {'HALLUCINATED': 1.3255301121249794, 'FAITHFUL': -0.5046869936399162, 'n': 'bfsd'}, {'length': 200}]
d[ 0.7 ] = [{'LABEL_0': 1.0334709184616804, 'n': 'grs'}, {'HALLUCINATED': 1.3255301121249794, 'FAITHFUL': -0.5046869936399162, 'n': 'bfsd'}, {'length': 200}]
d[ 0.8 ] = [{'LABEL_0': 1.0334709184616804, 'n': 'grs'}, {'HALLUCINATED': 1.3255301121249794, 'FAITHFUL': -0.5046869936399162, 'n': 'bfsd'}, {'length': 200}]
d[ 0.9 ] = [{'LABEL_0': 1.0334709184616804, 'n': 'grs'}, {'HALLUCINATED': 1.3255301121249794, 'FAITHFUL': -0.5046869936399162, 'n': 'bfsd'}, {'length': 200}]
d[ 1.0 ] = [{'LABEL_0': 1.0334709184616804, 'n': 'grs'}, {'HALLUCINATED': 1.3255301121249794, 'FAITHFUL': -0.5046869936399162, 'n': 'bfsd'}, {'length': 200}]
May I kindly ask for hints about the reason for these results? They get the same scores with different objective weights. Was it caused by not tuning the number of epochs for fine-tuning?
Kind regards,
Ethan
A declarative, efficient, and flexible JavaScript library for building user interfaces.
๐ Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. ๐๐๐
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google โค๏ธ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.