Project description

SuperAlignment package

This is the official github repo of pypi package SuperAlignment (https://pypi.org/project/SuperAlignment).

Motivated by OpenAI's alignment paper "WEAK-TO-STRONG GENERALIZATION: ELICITING STRONG CAPABILITIES WITH WEAK SUPERVISION"(https://openai.com/index/weak-to-strong-generalization/), we feel that it's a very interesting direction to define some common interface to train/eval superalignment models in this new area. Right now it has just started from scratch and you are very welcome to contact us if you have any cool ideas that would like to participate and collabratively commit to this repo.

pip install SuperAlignment

import SuperAlignment as sa

input_dict = {"text": "SuperAlignment"}

res = sa.api(input_dict, model=None, api_name="ArxivPaperAPI", start=0, max_results = 3)
paper_list = json.loads(res["text"])
print ("###### Text to Image Recent Paper List:")
for (i, paper_json) in enumerate(paper_list):
    print ("### PAPER %d" % (i+1))
    print (paper_json)

Common Interface of SuperAlignment Application


class YourSuperAlignmentAPI(BaseAPI):
    """docstring for ClassName"""
    def __init__(self, configs):
        super(YourSuperAlignmentAPI, self).__init__(configs)
        self.name = "xxxxx"

    def api(self, input_dict, model, kwargs):
        """
            Args:
                input_dict: dict, multi-modal input text, image, audio and video
                model: huggingface model of tf or pytoch
                kwargs: key-value args
            Return:
                res_dict: dict, multi-modal text text, image, audio and video
        """
        res_dict={}
        try:
            input_text = input_dict["text"]   # str
            input_image = input_dict["image"] # image path
            input_audio = input_dict["audio"] # audio path
            input_video = input_dict["video"] # video path

            res_dict["text"] = None
            res_dict["image"] = None
            res_dict["audio"] = None
            res_dict["video"] = None
        except Exception as e:
            print (e)
        return res_dict

SuperAlignment Losses

import torch
import torch.nn as nn
import torch.nn.functional as F

def auxConfidentLoss(alpha, t, x_weak, x_strong):
    """
        AUXILIARY CONFIDENCE LOSS as in weak to strong generalization paper
    """
    ce_loss = nn.CrossEntropyLoss()
    x_strong_ind = torch.relu(x_strong - t)
    aux_loss = alpha * ce_loss(x_strong, x_weak) + (1-alpha) * ce_loss(x_strong, x_strong_ind)
    return aux_loss

Awesome SuperAlignment Papers and Projects

2024 SuperAlignment Paper

PAPER	URL
WEAK-TO-STRONG GENERALIZATION: ELICITING STRONG CAPABILITIES WITH WEAK SUPERVISION	https://cdn.openai.com/papers/weak-to-strong-generalization.pdf
Strong and weak alignment of large language models with human values	https://arxiv.org/pdf/2408.04655
SUPER(FICIAL)-ALIGNMENT: STRONG MODELS MAY DECEIVE WEAK MODELS IN WEAK-TO-STRONG GENERALIZATION	https://arxiv.org/pdf/2406.11431
SELF-PLAY WITH EXECUTION FEEDBACK: IMPROVING INSTRUCTION-FOLLOWING CAPABILITIES OF LARGE LANGUAGE MODELS	https://arxiv.org/pdf/2406.13542
Quantifying the Gain in Weak-to-Strong Generalization	https://arxiv.org/abs/2405.15116
A Superalignment Framework in Autonomous Driving with Large Language Models	https://arxiv.org/abs/2406.05651
A Moral Imperative: The Need for Continual Superalignment of Large Language Models	https://arxiv.org/abs/2403.14683
Vision Superalignment: Weak-to-Strong Generalization for Vision Foundation Models	https://arxiv.org/abs/2402.03749
Improving Weak-to-Strong Generalization with Scalable Oversight and Ensemble Learning	https://arxiv.org/abs/2402.00667

AI Services Reviews and Ratings

Project details

These details have not been verified by PyPI

Project links

Release history Release notifications | RSS feed

This version

0.0.2

Oct 12, 2024

0.0.1

Aug 14, 2024

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distributions

No source distribution files available for this release.See tutorial on generating distribution archives.

Built Distribution

SuperAlignment-0.0.2-py3-none-any.whl (6.1 kB view details)

Uploaded Oct 12, 2024 Python 3

File details

Details for the file SuperAlignment-0.0.2-py3-none-any.whl.

File metadata

Download URL: SuperAlignment-0.0.2-py3-none-any.whl
Upload date: Oct 12, 2024
Size: 6.1 kB
Tags: Python 3
Uploaded using Trusted Publishing? No
Uploaded via: twine/5.1.1 CPython/3.11.2

File hashes

Hashes for SuperAlignment-0.0.2-py3-none-any.whl
Algorithm	Hash digest
SHA256	`883652064c29451cbd91c6ba01e8fdd45da38fba3ac9dca489d45d58fe17736d`
MD5	`18c0df405fa16912917a9637c94c2815`
BLAKE2b-256	`c5bea2b71d6dd3723f957e7373dd3054943ffb9f4cc344bf05f4561c21244cee`