yoavg

Reinforcement Learning for Language Models

Yoav Goldberg, April 2023.

Why RL?

With the release of the ChatGPT model and followup large language models (LLMs), there was a lot of discussion of the importance of "RLHF training", that is, "reinforcement learning from human feedback". I was puzzled for a while as to why RL (Reinforcement Learning) is better than learning from demonstrations (a.k.a supervised learning) for training language models. Shouldn't learning from demonstrations (or, in language model terminology "instruction fine tuning", learning to immitate human written answers) be sufficient? I came up with a theoretical argument that was somewhat convincing. But I came to realize there is an additional argumment which not only supports the case of RL training, but also requires it, in particular for models like ChatGPT. This additional argument is spelled out in (the first half of) a talk by John Schulman from OpenAI. This post pretty much

efonte

import re


strings = set()

with open('texts.txt', mode='r', encoding='utf-8') as file_input:
    content = file_input.read()
    regExStr = r'^\s*\d string title = "(?:tooltip\d+|Name|Dialogue Text)"\n\s*\d string value = "((?!(START|input|\w+\(\)|\!\(\w+\(\)\))).+)"$'
    compiled = re.compile(regExStr, re.MULTILINE)
    matched = compiled.finditer(content)

stephenroller

#!/usr/bin/env python3

"""
Example of a generic Mixout implementation. (Lee et al., 2019).
https://arxiv.org/abs/1909.11299

Implementation by Stephen Roller (https://stephenroller.com).

Updated 2020-02-10 to include 1/(1 - p) correction term. Thanks to
Cheolhyoung Lee for making this correction.

yzh119

from __future__ import print_function
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable

def sample_gumbel(shape, eps=1e-20):
    U = torch.rand(shape).cuda()
    return -Variable(torch.log(-torch.log(U + eps) + eps))

BaseCase

Deconstruct 2017 Bibliography

Here are all of the resources mentioned by Deconstruct 2017 speakers, along with who recommended what. Please post a comment if I missed something or have an error!

DC 2017 Speakers' Choice Gold Medalist

• Seeing Like a State by James Scott

Books

• Public Opinion by Walter Lippmann (Evan Czaplicki)
• A Pattern Language by Christopher Alexander (Brian Marick)
• Domain Driven Design by Eric Evans (Brian Marick)

kastnerkyle

"""
TreeLSTM[1] implementation in Pytorch

Based on dynet benchmarks : 
  https://github.com/neulab/dynet-benchmark/blob/master/dynet-py/treenn.py
  https://github.com/neulab/dynet-benchmark/blob/master/chainer/treenn.py
Other References:
  https://github.com/pytorch/examples/tree/master/word_language_model
  https://github.com/pfnet/chainer/blob/29c67fe1f2140fa8637201505b4c5e8556fad809/chainer/functions/activation/slstm.py
  https://github.com/stanfordnlp/treelstm

Kronopath

Speaker: Mike Walfish

mwalfish@nyu.edu

Reccomended reading: http://www.eecs.harvard.edu/htk/phdadvice/

Before

Should you do one?

maryrosecook

We couldn’t find that file to show.

coolaj86

Please petition Github to support HTTPS on github pages: https://github.com/contact

Here's what I wrote:

Obviously, a lot of people want HTTPS for github pages:

• isaacs/github#156

Until recently, that would be difficult to implement but, as it turns out, the implementation is pretty much complete:

• https://letsencrypt.org >* https://github.com/letsencrypt/lets-encrypt-preview

bcomnes

# Fix agent forwarding
# https://gist.github.com/martijnvermaat/8070533
# http://techblog.appnexus.com/2011/managing-ssh-sockets-in-gnu-screen/
# See .ssh/rc for socket linking
unsetenv SSH_AUTH_SOCK
setenv SSH_AUTH_SOCK $HOME/.ssh/ssh_auth_sock.$HOSTNAME