Report for milestone submission - just dont try to compile, the pdf i…

…s from overleaf where my bibtex actually works

report/references.bib

@@ -1,16 +1,71 @@
-@book{aho1972theory,
-    author  = {Alfred V. Aho and Jeffrey D. Ullman},
-    title   = {The Theory of Parsing, Translation and Compiling},
-    year    = "1972",
-    volume  = "1",
-    publisher = {Prentice-Hall},
-    address = {Englewood Cliffs, NJ}
+@misc{hendrycks2023aligning,
+      title={Aligning AI With Shared Human Values}, 
+      author={Dan Hendrycks and Collin Burns and Steven Basart and Andrew Critch and Jerry Li and Dawn Song and Jacob Steinhardt},
+      year={2023},
+      eprint={2008.02275},
+      archivePrefix={arXiv},
+      primaryClass={cs.CY}
 }
 
-@inproceedings{andrew2007scalable,
-  title={Scalable training of {L1}-regularized log-linear models},
-  author={Andrew, Galen and Gao, Jianfeng},
-  booktitle={Proceedings of the 24th International Conference on Machine Learning},
-  pages={33--40},
-  year={2007},
+@misc{liu2019roberta,
+      title={RoBERTa: A Robustly Optimized BERT Pretraining Approach}, 
+      author={Yinhan Liu and Myle Ott and Naman Goyal and Jingfei Du and Mandar Joshi and Danqi Chen and Omer Levy and Mike Lewis and Luke Zettlemoyer and Veselin Stoyanov},
+      year={2019},
+      eprint={1907.11692},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
 }
+
+@misc{lourie2021scruples,
+      title={Scruples: A Corpus of Community Ethical Judgments on 32,000 Real-Life Anecdotes}, 
+      author={Nicholas Lourie and Ronan Le Bras and Yejin Choi},
+      year={2021},
+      eprint={2008.09094},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+
+@misc{ziems2022moral,
+      title={The Moral Integrity Corpus: A Benchmark for Ethical Dialogue Systems}, 
+      author={Caleb Ziems and Jane A. Yu and Yi-Chia Wang and Alon Halevy and Diyi Yang},
+      year={2022},
+      eprint={2204.03021},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+
+@misc{touvron2023llama,
+      title={Llama 2: Open Foundation and Fine-Tuned Chat Models}, 
+      author={Hugo Touvron and Louis Martin and Kevin Stone and Peter Albert and Amjad Almahairi and Yasmine Babaei and Nikolay Bashlykov and Soumya Batra and Prajjwal Bhargava and Shruti Bhosale and Dan Bikel and Lukas Blecher and Cristian Canton Ferrer and Moya Chen and Guillem Cucurull and David Esiobu and Jude Fernandes and Jeremy Fu and Wenyin Fu and Brian Fuller and Cynthia Gao and Vedanuj Goswami and Naman Goyal and Anthony Hartshorn and Saghar Hosseini and Rui Hou and Hakan Inan and Marcin Kardas and Viktor Kerkez and Madian Khabsa and Isabel Kloumann and Artem Korenev and Punit Singh Koura and Marie-Anne Lachaux and Thibaut Lavril and Jenya Lee and Diana Liskovich and Yinghai Lu and Yuning Mao and Xavier Martinet and Todor Mihaylov and Pushkar Mishra and Igor Molybog and Yixin Nie and Andrew Poulton and Jeremy Reizenstein and Rashi Rungta and Kalyan Saladi and Alan Schelten and Ruan Silva and Eric Michael Smith and Ranjan Subramanian and Xiaoqing Ellen Tan and Binh Tang and Ross Taylor and Adina Williams and Jian Xiang Kuan and Puxin Xu and Zheng Yan and Iliyan Zarov and Yuchen Zhang and Angela Fan and Melanie Kambadur and Sharan Narang and Aurelien Rodriguez and Robert Stojnic and Sergey Edunov and Thomas Scialom},
+      year={2023},
+      eprint={2307.09288},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+
+@misc{lan2020albert,
+      title={ALBERT: A Lite BERT for Self-supervised Learning of Language Representations}, 
+      author={Zhenzhong Lan and Mingda Chen and Sebastian Goodman and Kevin Gimpel and Piyush Sharma and Radu Soricut},
+      year={2020},
+      eprint={1909.11942},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+
+@misc{sanh2020distilbert,
+      title={DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter}, 
+      author={Victor Sanh and Lysandre Debut and Julien Chaumond and Thomas Wolf},
+      year={2020},
+      eprint={1910.01108},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+
+@misc{brown2020language,
+      title={Language Models are Few-Shot Learners}, 
+      author={Tom B. Brown and Benjamin Mann and Nick Ryder and Melanie Subbiah and Jared Kaplan and Prafulla Dhariwal and Arvind Neelakantan and Pranav Shyam and Girish Sastry and Amanda Askell and Sandhini Agarwal and Ariel Herbert-Voss and Gretchen Krueger and Tom Henighan and Rewon Child and Aditya Ramesh and Daniel M. Ziegler and Jeffrey Wu and Clemens Winter and Christopher Hesse and Mark Chen and Eric Sigler and Mateusz Litwin and Scott Gray and Benjamin Chess and Jack Clark and Christopher Berner and Sam McCandlish and Alec Radford and Ilya Sutskever and Dario Amodei},
+      year={2020},
+      eprint={2005.14165},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}

report/report.tex

@@ -27,85 +27,143 @@
 
 \newcommand\BibTeX{B\textsc{ib}\TeX}
 
-\title{Classifying news articles by partisan lean}
+\title{Digital Socrates: Classifying Action Statements as (Un)Ethical}
 
-\author{Alice Cooper \\
-  School's Out University / Detriot, MI \\
-  \texttt{email@domain} \\\And
-  Bob Seger \\
-  Night Moves U / Detroit, MI \\
-  \texttt{email@domain} \\}
+\author{Harper Lyon \\
+  Tulane University / New Orleans, LA \\
+  \texttt{[email protected]}}
 
 \date{}
 
 \begin{document}
 \maketitle
-\begin{abstract}
-Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
-\end{abstract}
 
+\section{Problem Overview}
 
-\section{Introduction}
+The idea of ethical machines is attractive, but challenging to implement in practice. One approach is to borrow techniques from sentiment analysis \cite{hendrycks2023aligning} to classify everyday descriptions of actions as ethical or not, creating ideally a system which can give human-like responses to human questions about decisions. This problem contains many of the same inherent challenges as sentiment analysis, as subtle differences in language use can shift the meaning of documents massively - consider the difference between "I will not break my promise even for money" and "I will not break my promise except for even more money" - but also in that it seemingly requires more modeling of meaning than simple sentiment analysis, since ethical judgements are inherently tied to actions in a way that movie reviews or teaching evaluations are not. 
 
-This is a skeleton latex document for you to write your report.
+The "hook" for my project is the creation of a webpage where a digital facsimile of Socrates, the famously wise and annoying Athenian, tells you whether or not what you are planning to do is ethical, but for that I will need a classification model capable of accurately making that decision.
 
+\section{Data}
 
-Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
-
-\section{Related Work}
-
-Citations within the text appear in parentheses as~\citep{aho1972theory} or, if the author's name appears in the text itself, as \citet{andrew2007scalable}.
+My primary data source is the ETHICS \cite{hendrycks2023aligning} dataset, which is a dataset combining five semi-distinct text classification tasks related to various areas of ethical reasoning. For my purposes, only two subsections of the dataset are useful, and these are as follows:
 
+\begin{table}[htbp]
+    \centering
+    \caption{Dataset Divisions}
+    \begin{tabular}{llll}
+        \toprule
+        & \textbf{Rows} & \textbf{Classes}\\
+        \midrule
+        Commonsense & $21.8$k & $0:11.49$k, $1:10.26$k\\
+        Justice & $26.5$k & $0:12.31$k, $1:14.23$k\\
+        \bottomrule
+    \end{tabular}
+    \label{tab:my_table}
+\end{table}
 
+These two subsets are similar, but not exactly the same. The commonsense dataset consists of two types of document class pairs, the first being straightforward action descriptions 
+like ("I left her bleeding on the snowy hillside.", 1) and the other being posts sourced from reddit.com/r/AITAH, an internet forum where users post stories from their lives and receive crowdsourced moral judgements. We can question of the wisdom of relying on these judgements, but they are a convinient corpus for this project. 
 
-Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
+The justice dataset consists of statements of the for I (deserve/did) \textit{action} because \textit{reason}, for example ("I deserve to get a nice haircut from my barber because I paid him to make my hair look nice.", 1). Some work is required to reconcile these datasets to a single task - mainly switching the labels to match - I see no reason that they cannot be accomplished by the same model or at least by similar models trained in similar ways.
 
+I am also in the early stages of finding additional data sources to augment my training data, more information on those potential sources \cite{lourie2021scruples}\cite{ziems2022moral} in Related Work.
 
 \section{Methods}
 
-Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
+My goal is to evaluate three distinct approaches to this task at different "weights", in particular
 
+\begin{enumerate}
+    \item Logistic Regression
+    \item Finetuned BERT
+    \item Llama
+\end{enumerate}
 
+Logistic regression serves as a baseline model, since it will ultimately work through simple word inclusion, but it should give us a good idea, more or less, of how difficult the task is. Otherwise I don't expect it to be all that interesting or play a large part in my final application.
 
-\section{Results}
+Much of the existing work in this area focuses on BERT based models \cite{lourie2021scruples}\cite{hendrycks2023aligning}, so these are a natural place to begin the real work. There are many BERT varieties, so I expect to try several options (currently I have only worked with distilbert). However, the best BERT based performance so far outside of the commonsense data subset is moderate at best, with ALBERT\cite{lan2020albert} achieving $59.9\%$ on the justice task and $85.5\%$ on commonsense. I don't expect to beat those numbers given my limited resources and time, but it's clear that especially on justice BERT is not up to the task of top notch performance.
 
-Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
+What I really want to explore is using an LLM like Llama to generate classifications since this is an area left open by the originators of the ETHICS dataset. I plan to use Llama-2-7b since I am limited by computing power, and that is the most moderate of the new models available. I have already been given access by Meta to the weights for this model, so I will be beginning work here soon.
 
+Evaluation will be straightforward, I am interested in achieving high accuracy and not much else. If I can match the median results of the original paper I'll be happy with my outcomes. I'm also interested in examining the loss variance
 
-Table~\ref{tab:a_table} shows a table. Note that we refer to output generated by \texttt{Experiments.ipynb}. This way, whenever we re-run our notebook, we can regenerate the paper with the latest results.
+\section{Preliminary Results}
 
-\begin{table}[ht]
-\centering
-% note that we can refer to tables generated by our Experiments.ipynb notbook.
-\input{../notebooks/table.tex}
-\caption{\label{tab:a_table} A caption. }
-\end{table}
-
-Figure~\ref{fig:a_label} shows a figure
+So far I have (on the commonsense subset only) implemented the Logistic Regression model and made a first clumsy pass at finetuning a BERT variant called DistilBERT \cite{sanh2020distilbert}, which I chose for no other reason than that is quick to train as a proof of concept and seems to be the "standard" BERT variant used for classification absent other considerations. I've achieved extremely poor performance with both models, hitting exactly chance on Logistic Regression (recall that we have a 50/50 class split) for Logistic Regression \ref{fig:log_reg} and actually performing worse than chance with DistilBERT \ref{fig:distil}. I'm not surprised by the results of logistic regression, but I am disappointed at how poorly DistilBERT performed - there is definitely an over-fitting or analogous issue since the model acheived ~$85\%$ accuracy on the validation split.
 
-\begin{figure}[ht]
-	\centering
-	% note that we can refer to figures generated by our Experiments.ipynb notbook.
-	\includegraphics[width=3.5in]{../notebooks/results.pdf}
-	\caption{A caption}
-	\label{fig:a_label}
+\begin{figure}
+  \centering
+  \includegraphics[width=\columnwidth]{log_reg.png}
+  \caption{Logistic Regression Confusion Matrix}
+  \label{fig:log_reg}
 \end{figure}
 
-\section{Discussion}
-
-Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
-
-
-\section{Conclusion}
-Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
-
-
-\section{Division of Labor}
-Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
-
+\begin{figure}
+  \centering
+  \includegraphics[width=\columnwidth]{distil.png}
+  \caption{DistilBERT Confusion Matrix}
+  \label{fig:distil}
+\end{figure}
 
+\section{Related Work}
 
+\subsection{BERT Related Work}
+\vspace{.5em}
+
+\begin{itemize}
+    \item\textbf{Title:} Aligning AI With Shared Human Values \\
+    \textbf{Citation:} \cite{hendrycks2023aligning} \\
+    \textbf{Description:} This is the paper which originated the ETHICS dataset and inspired a lot of this project - to a large extent all I'm trying to do is extend their work slightly. They focus almost entirely on BERT based approaches, especially RoBERTa (see below), so I hope that by taking advantage of improvements in BERTology (as it where) and by introducing LLMs I can make some meaningful improvements to their work.
+
+    \item\textbf{Title:} Scruples: A Corpus of Community Ethical Judgments on 32,000 Real-Life Anecdotes \\
+    \textbf{Citation:} \cite{lourie2021scruples} \\
+    \textbf{Description:} This is a potential source of additional data, which is especially important since I seem to be having finetuning problems. These "Community Judgements" are also sourced from reddit.com/r/AMITAH, so they should dovetail well with the ETHICS dataset.
+
+    \item\textbf{Title:} RoBERTa: A Robustly Optimized BERT Pretraining Approach \\
+    \textbf{Citation:} \cite{liu2019roberta} \\
+    \textbf{Description:} This paper describes many of the finetuning techniques used by Hendrycks et al. to achieve their best performance on ETHICS, so I'm hoping that I can use RoBERTa, and their finetuning techniques more broadly, to improve my abysmall current DistilBERT accuracy.
+\end{itemize}
+
+\subsection{LLM Related Work}
+
+\begin{itemize}
+    \item\textbf{Title:} The Moral Integrity Corpus: A Benchmark for Ethical Dialogue Systems\\
+    \textbf{Citation:} \cite{ziems2022moral} \\
+    \textbf{Description:} I'll admit, this paper is more for me than for this project specifically. They propose a Seq-Seq moral reasoning process/prompt engineering technique which injects essentially attacks on the LLM's moral reasoning in the form of sentences like "You should not judge people negatively based on race" to test and improve output. I'm still not sure if any output beyond classification from Llama is going to play a role in my final product, but if so I do think there are some valuable insights here. 
+
+    \item\textbf{Title:} Language Models are Few-Shot Learners\\
+    \textbf{Citation:} \cite{brown2020language} \\
+    \textbf{Description:} This paper describes a variety of techniques for adapting LLM's Seq-Seq functioning for other tasks including text classification. This is more or less the paper on doing so, so it should be helpful as I move into the third stage of the project.
+\end{itemize}
+
+\section{Timeline}
+
+I plan to hold to the following rough timeline of tasks:
+
+\begin{table}[htbp]
+    \centering
+    \begin{tabular}{p{0.4\columnwidth} p{0.4\columnwidth}}
+        \toprule
+        \textbf{Task} & \textbf{Estimated Completion} \\
+        \midrule
+        Install Llama-2-7b, begin fine tuning & 3/29 \\
+        \midrule
+        Tune Additional BERT models & 4/4 \\
+        \midrule
+        Determine best model for web integration & 4/11 \\
+        \midrule
+        Implement "Digital Socrates" page & 4/16 \\
+        \midrule
+        Complete presentation & 4/20 \\
+        \midrule
+        Complete report & 4/28
+        \bottomrule
+    \end{tabular}
+    \caption{Timeline}
+    \label{tab:task_completion}
+\end{table}
 
+With some expected slop and slippage here and there.
 
 \bibliography{references}
 \bibliographystyle{acl_natbib}