Jeremy Howard jph00

https://wandb.ai/rom1504/dalle2_train_decoder/runs/mic5buox/files/decoder_config.json

get dalle2

get the config file

get these 2 .sh

run sbatch start_big.sh

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

	#!/usr/bin/env bash

	# Install miniconda - skip if you've already got conda
	cd ~
	mkdir -p downloads
	cd downloads
	wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
	bash Miniconda3-latest-Linux-x86_64.sh -b
	conda init bash
	cd ..

	<opml version="1.0">
	<head>
	<dateCreated>Mon, 21 Nov 2022 20:41:45 +0000</dateCreated>
	</head>
	<body>
	<outline text="FastML" title="FastML" type="rss" xmlUrl="http://fastml.com/atom.xml" htmlUrl="http://fastml.com/"/>
	<outline text="colah's blog" title="colah's blog" type="rss" xmlUrl="http://colah.github.io/rss" htmlUrl="http://colah.github.io/"/>
	<outline text="Math ∩ Programming" title="Math ∩ Programming" type="rss" xmlUrl="https://jeremykun.com/feed/" htmlUrl="https://jeremykun.com"/>
	<outline text="Pete Warden's blog" title="Pete Warden's blog" type="rss" xmlUrl="https://petewarden.com/feed/" htmlUrl="https://petewarden.com"/>
	<outline text="Two Minute Papers" title="Two Minute Papers" type="rss" xmlUrl="https://www.youtube.com/feeds/videos.xml?channel_id=UCbfYPyITQ-7l4upoX8nvctg" htmlUrl="https://www.youtube.com/channel/UCbfYPyITQ-7l4upoX8nvctg"/>

	# Credit 🙏: I just used the example from langchain docs and it works quite well: https://python.langchain.com/en/latest/use_cases/question_answering.html
	# Note 2: The Arxiv -> PDF logic is a bit messy, I'm sure it can be done better
	# Note 3: Please install the following:

	# To run:

	# Save this in a `app.py`
	# pip install arxiv PyPDF2 langchain chromadb

	# The chat feature was shipped in H2O nightly this week, we will need to install from nightly link:

	import torch
	from datasets import load_dataset
	import argparse
	import os
	import math
	from itertools import chain
	from datetime import timedelta
	from torch.utils.data import DataLoader
	from accelerate import Accelerator
	from accelerate.utils import (DummyOptim, DummyScheduler,

	This table contains data from multiple software versions. Some hyperparamter names are "NaN" meaning, they did not exist in that software version. The best 7B result is 40.08 MMLU.

	================================================================================
	Config: learning_rate: 0.005, adam_beta2: 0.999, lora_dropout: 0.0 , max_grad_norm: 1.0 , max_steps: 7320, lr_scheduler_type: <SchedulerType.COSINE: cosine>, weight_decay: 0.0 , base_model: /gscratch/zlab/llama/7B, quant_type: nf4 , gradient_accumulation_steps: 6 , per_device_train_batch_size: 2
	acc mean (SE): 0.2290 (nan). 95% CI (nan, nan). Sample size: 1
	================================================================================
	Config: learning_rate: 0.0002, adam_beta2: 0.999, lora_dropout: 0.0 , max_grad_norm: 0.3 , max_steps: 9750, lr_scheduler_type: <SchedulerType.CONSTANT: constant>, weight_decay: 0.0 , base_model: NaN , quant_type: nf4 , gradient_accumulation_steps: 2 , per_device_train_batch_size: 8
	acc mean (SE): 0.2290 (0.0

	Config: max_steps: 17500, lora_r: 64 , lr: 0.0002, bf16: False, lora_modules: all , bits: 4 , full_finetune: False, lora_dropout: 0.0 , warmup_steps: 100 , compress_statistics: True, dataset: NaN , gradient_accumulation_steps: NaN , learning_rate: NaN , quant_type: fp4 , adam_beta2: 0.999, update_freq: 6
	eval_bert_f1 mean (SE): 64.8716 (21.8331). 95% CI (22.079, 107.664). Sample size: 4
	eval_bert_f1 mean (SE): 64.8716 (21.8331). 95% CI (22.079, 107.664). Sample size: 4
	eval_rougeL mean (SE): 33.1083 (19.1162). 95% CI (-4.359, 70.576). Sample size: 4
	================================================================================
	Config: max_steps: 17500, lora_r: 64 , lr: 0.0002, bf16: False, lora_modules: all , bits: 4 , full_finetune: False, lora_dropout: 0.0 , warmup_steps: 100 , compress_statistics: False, dataset: NaN , gradient_accumulation_steps: NaN , learning_rate: NaN , quant_type: fp4 , adam_beta2: 0.999, update_freq: 6
	eval_bert_f1 mean (SE): 67.0044 (22.3593). 95% CI (23.180, 110.829).

	# coding=utf-8
	# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
	#
	# Licensed under the Apache License, Version 2.0 (the "License");
	# you may not use this file except in compliance with the License.
	# You may obtain a copy of the License at
	#
	# http://www.apache.org/licenses/LICENSE-2.0
	#
	# Unless required by applicable law or agreed to in writing, software