Shreyansh Singh shreyansh26

Redirects HF paper pages to arXiv.

Install Redirector

Chrome: https://chrome.google.com/webstore/detail/redirector/ocgpenflpmgnfapjedencafcfakcekcd

Firefox: https://addons.mozilla.org/en-US/firefox/addon/redirector/

Add redirect

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

	import torch.nn as nn
	import copy
	import torch
	from torch.nn.attention.flex_attention import flex_attention, create_block_mask, or_masks, create_mask
	from triton.testing import do_bench
	from functools import partial

	torch.set_default_device('cuda')

	B = 4

	import torch
	import os
	import json
	from safetensors.torch import load_file, save_file

	def replicate_lora_a(name: str, weight: "torch.Tensor") -> dict[str, "torch.Tensor"]:
	prefix, suffix = name.split('qkv_proj')
	res = {}
	for t in ['q_proj', 'k_proj', 'v_proj']:
	name = f"{prefix}{t}{suffix}"

	import torch
	torch.set_default_device('cuda')
	from triton.testing import do_bench
	from collections import defaultdict
	from functools import partial
	import random
	random.seed(0)

	def get_flops(A, B):
	ms = do_bench(lambda: torch.mm(A, B))

	#!/usr/bin/env python

	import argparse
	import torch
	from transformers import GPTJForCausalLM, GPTJConfig
	# Note: these need the git version of Transformers as of 7/22/2022
	from transformers import CodeGenTokenizer, CodeGenForCausalLM
	from transformers import CODEGEN_PRETRAINED_MODEL_ARCHIVE_LIST

	parser = argparse.ArgumentParser('Convert SalesForce CodeGen model to GPT-J')

	# http://editorconfig.org/#file-format-details
	root = true

	[*]
	charset = utf-8
	end_of_line = lf
	indent_size = 4
	indent_style = space
	insert_final_newline = true
	trim_trailing_whitespace = true

	# This isn't supposed to run as a bash script, i named it with ".sh" for syntax highlighting.

	# https://developer.nvidia.com/nsight-systems
	# https://docs.nvidia.com/nsight-systems/profiling/index.html

	# My preferred nsys (command line executable used to create profiles) commands
	#
	# In your script, write
	# torch.cuda.nvtx.range_push("region name")
	# ...

	import math
	import torch
	import torch.nn as nn
	from torch.nn import functional as F


	class RelativePositionBias(nn.Module):
	def __init__(self, bidirectional=True, num_buckets=32, max_distance=128, n_heads=2):
	super(RelativePositionBias, self).__init__()
	self.bidirectional = bidirectional