Tokarev Igor tokarev-i-v

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

Purpose

Bootstrap knowledge of LLMs ASAP. With a bias/focus to GPT.

Avoid being a link dump. Try to provide only valuable well tuned information.

Prelude

Neural network links before starting with transformers.

	# train_grpo.py
	#
	# See https://github.com/willccbb/verifiers for ongoing developments
	#
	import re
	import torch
	from datasets import load_dataset, Dataset
	from transformers import AutoTokenizer, AutoModelForCausalLM
	from peft import LoraConfig
	from trl import GRPOConfig, GRPOTrainer

	def plot_weights(model, layer_num, single_channel = True, collated = False):

	#extracting the model features at the particular layer number
	layer = model.features[layer_num]

	#checking whether the layer is convolution layer or not
	if isinstance(layer, nn.Conv2d):
	#getting the weight tensor data
	weight_tensor = model.features[layer_num].weight.data

	from fabric.api import env, run as fabric_run
	from fabric.contrib.project import rsync_project

	env.app_path = '/home/backend'
	env.user = 'circle'

	env.use_ssh_config = True
	env.disable_knodwn_hosts = True
	env.colorize_errors = True

	{
	"name": "webpack-sass",
	"version": "1.0.0",
	"scripts": {
	"start": "webpack-dev-server --open --mode development",
	"build": "webpack -p"
	},
	"devDependencies": {
	"babel-core": "^6.26.0",
	"babel-loader": "^7.1.4",

	/**
	* @author richt / http://richt.me
	* @author WestLangley / http://github.com/WestLangley
	*
	* W3C Device Orientation control (http://w3c.github.io/deviceorientation/spec-source-orientation.html)
	*/

	THREE.DeviceOrientationControls = function( object ) {

	var scope = this;

	<!DOCTYPE html>
	<html>
	<head>
	<title>Real-time LaTeX in browser</title>
	<!-- https://github.com/mathjax/MathJax/blob/master/test/sample-dynamic-2.html -->
	<!-- Copyright (c) 2012-2015 The MathJax Consortium -->
	<meta http-equiv="Content-Type" content="text/html; charset=UTF-8" />
	<meta http-equiv="X-UA-Compatible" content="IE=edge" />
	<meta name="viewport" content="width=device-width, initial-scale=1">

	curl --include \
	--no-buffer \
	--header "Connection: Upgrade" \
	--header "Upgrade: websocket" \
	--header "Host: example.com:80" \
	--header "Origin: http://example.com:80" \
	--header "Sec-WebSocket-Key: SGVsbG8sIHdvcmxkIQ==" \
	--header "Sec-WebSocket-Version: 13" \
	http://example.com:80/

	#!/usr/bin/env python
	# -- coding: utf-8 --
	#
	# A5.py
	#
	# Author: overxfl0w13
	#

	def header():
	print """