llhe

|Title of the paper |Project cited |Published media |Download URL

llhe

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

llhe

[
  {
    "id": "camera0",
    "parentIds": [],
    "status": "running"
  },
  {
    "id": "lidar0",
    "parentIds": [],
    "status": "running"

llhe

name: "MobileNet-SSD"
input: "data"
input_shape {
  dim: 1
  dim: 3
  dim: 300
  dim: 300
}
layer {
  name: "conv0"

llhe

• Color banding
• Purple fringing
• Lens flare
• Ringing artifacts
• Posterization
• Aliasing/Moiré Pattern
• Chromatic aberration
• Rolling shutter
• Vignetting
• Noise

llhe

model_name                     device_name   soc        abi           runtime   init       warmup     run_avg    tuned
mobilenet_v2                   polaris       sdm845     armeabi-v7a   GPU       42.868     11.087     9.908      True
mobilenet_v2                   MI MAX        msm8952    armeabi-v7a   GPU       122.791    43.038     39.875     True
mobilenet_v2                   BKL-AL00      kirin970   armeabi-v7a   GPU       767.932    1226.373   47.597     True
mobilenet_v2                   polaris       sdm845     arm64-v8a     GPU       42.3       10.737     10.004     True
mobilenet_v2                   MI MAX        msm8952    arm64-v8a     GPU       129.123    42.584     39.552     True
mobilenet_v2                   BKL-AL00      kirin970   arm64-v8a     GPU       753.43     1170.291   48.016     True
mobilenet_v2                   polaris       sdm845     armeabi-v7a   CPU       16.035     69.761     41.627     False
mobilenet_v2                   MI MAX        msm8952    armeabi-v7a   CPU     

llhe

set nu
set ruler
set tabstop=2 shiftwidth=2 expandtab
set colorcolumn=80

" pathogen
execute pathogen#infect()
syntax on
filetype plugin indent on

llhe

/*
 * g++ -std=c++11 -pthread memtest.cc
 */

#include <cstring>

#include <chrono>
#include <condition_variable>
#include <functional>
#include <future>

llhe

"""Benchmark tensorflow distributed by assigning a tensor between two workers.

Usage:
  Start worker 1:
  python rdma_bench.py --workers="hostname1:port,hostname2:port" --protocol=grpc+verbs --task 0

  Start worker 2:
  python rdma_bench.py --workers="hostname1:port,hostname2:port" --protocol=grpc+verbs --task 1

  Run the tests:

llhe

import argparse
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data

def train():
  mnist = input_data.read_data_sets("/tmp/data",
                                    one_hot=True,
                                    fake_data=False)
  sess = tf.InteractiveSession()