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

laggardkernel

Comparison of ZSH frameworks and plugin managers

Changelog

• update 1: add a FAQ section
• update 2: benchmark chart and feature comparison table
• update 3:
  • improve the table with missing features for antigen
  • new zplg times result

TLDR

seanjtaylor

carlislerainey

# code to understand product terms and probit models

# clear workspace
rm(list = ls())

# load packages
library(tidyverse)
library(plotly)

# parameters

benmarwick

# metadata
data_source <- "https://docs.google.com/spreadsheets/d/1xa0iLqYKz8x9Yc_rfhtmSOJQ2EGgeUVjvV4A8LsIaxY/htmlview?sle=true#gid=0"
data_collectors <- "Jeremy Pressman (@djpressman, U of Connecticut) and\nErica Chenoweth (@EricaChenoweth, U of Denver)"
code_gist <- "https://gist.github.com/benmarwick/a1ac9c7235ebef542824512162ff2f44"


#  ------------------------------------------------------------------------

# read in data from google sheets to get a data frame

mattblackwell

group.mean <- function(x, group) {
  out <- tapply(x, group, mean, na.rm = TRUE)
  out[group]
}

data(state.fips, package = "maps")
state.fips <- unique(state.fips[,c("fips","abb")])
state.fips$abb <- as.character(state.fips$abb)
state.fips <- rbind(state.fips, c(2, "AK"))
state.fips <- rbind(state.fips, c(15, "HI"))

mbjoseph

# Simulation script for factor analysis ala Leung & Drton (2016) ----------
library(rstan)
library(bayesplot)

m <- 5 # dimension of observed data (e.g., # traits)
k <- 2 # number of latent factors
n <- 100 # number of sample units (e.g., # species)

# residual variance matrix (is diagonal)
Omega <- diag(.3 + abs(rnorm(m, sd = .3)))

leeper

# Check URLs in a document

## This code will extract URLs from a text document using regex,
## then execute an HTTP HEAD request on each and report whether
## the request failed, whether a redirect occurred, etc. It might
## be useful for cleaning up linkrot.

if (!require("httr")) {
    install.packages("httr", repos = "http://cran.rstudio.com/")
}

hrbrmstr

bishboria

These links no longer work. Springer have pulled the free plug.

Graduate texts in mathematics

duplicates = multiple editions

A Classical Introduction to Modern Number Theory, Kenneth Ireland Michael Rosen

A Classical Introduction to Modern Number Theory, Kenneth Ireland Michael Rosen