Merge pull request #925 from anvichip/main

Blog Post - 2: Autograder_Anvi Kohli

Author: slieggi

content/report/osre25/ucsc/autograder/20250725-anvichip/index.md

@@ -0,0 +1,76 @@
+---
+title: "Midway Report: LINQS - Autograder (LLM Detector)"
+subtitle: "Halfway through the work!"
+summary: "Midterm progress report on my GSoC'25 project with OSPO."
+authors:
+  - anvichip
+tags: ["osre25"]
+categories: ["Artificial Intelligence","Machine Learning", "LLMs"]
+date: 2025-07-31
+lastmod: 2025-07-31
+featured: true
+draft: false
+---
+
+Hello everyone, I’m Anvi Kohli and in this blog post, I’ll be sharing my journey as a GSoC contributor. 
+This summer I am contributing to the [LINQS Autograder project](https://ucsc-ospo.github.io/project/osre25/ucsc/autograder/) under the mentorship of Eriq Augustine and Lucas Ellenberger. 
+The goal of my project is to build a tool that can detect code generated with AI. 
+You can read my proposal here: [Proposal](https://summerofcode.withgoogle.com/programs/2025/projects/jxBUpvoM).
+My first blog is available here: [Blog Post 1](https://ucsc-ospo.github.io/report/osre25/ucsc/autograder/).
+
+## Project Overview
+
+The Autograder Server is an open source project that automatically grades programming assignments in real-time. 
+The project includes support for a variety of programming languages for evaluation. 
+Due to the rise of tools like GitHub Copilot and ChatGPT, students are increasingly relying on AI to complete their coding assignments. 
+This poses a problem as it becomes challenging to uphold fairness in grading and ensure that students are learning.  
+Our project aims to address this issue by creating a system that provides a confidence score to indicate that a piece of code was written by an AI tool.
+
+## Progress, Challenges, & Learnings
+
+### Exploration of Existing Tools and Systems
+
+My mentor, Eriq Augustine, advised me to begin with simpler methodologies before progressing to more complex ones. 
+There are a several possible approaches to detect AI generated code including training models from scratch, designing custom detection algorithms, and using and adapting existing open-source tools. 
+Since training a model from scratch requires an enormous amount of training data to be curated first, in the interest of time, we chose to begin by exploring pre-existing solutions and evaluating their performance.
+So first off, I conducted an in-depth exploration of open-source repositories that detect AI-generated code. 
+By building upon existing open source solutions, we can focus on enhancing the capabilities of pre-built tools and fine-tuning models.
+Training these models on large, diverse datasets can make the models more accurate, robust, and adaptable.
+
+Exploring open source solutions helped me gain an understanding of the current work and ongoing efforts in the detection of AI-generated code. 
+This exploration helped me identify the gaps that remain in the current tools and where there's room for improvement.
+
+### Transfer Learning
+
+While exploring existing tools and research papers, I found that transfer learning has shown promising results in the detection of AI-generated code. 
+For example, many studies fine-tuned pre-trained models like CodeBERT on labeled datasets containing AI and human-written code.
+Building on this, I curated a collection of publicly available datasets that could be used for this purpose.
+However, during this process, I noticed that open-source, relevant, and good quality datasets are limited. 
+They also vary widely in format, language coverage, and overall quality. 
+Some focus on a single programming language, while others span multiple languages. 
+Often, these datasets also lack sufficient examples. 
+By standardizing these disorganized resources, we can create a comprehensive, multi-language dataset suitable for AI code detection. 
+
+Currently, I’m working on fine-tuning these models using the open source datasets and evaluating their effectiveness in classifying AI-generated from human written code.
+
+### Contributing to Autograder Repository
+
+To help me get familiar with our existing codebase and gain hands-on experience with the Go programming language, my mentor assigned me an [Open Issue](https://github.com/edulinq/autograder-server/issues/141) of the repo. 
+Here is my progress on the same: [PR#194](https://github.com/edulinq/autograder-server/pull/194).
+
+As mentioned, the Autograder is a tool used to evaluate programming assignments.
+One of the features of the autograder server is it's ability to provide code analysis across a large number of code submissions.
+It leverages source code plagiarism detection engines like [JPlag](https://helmholtz.software/software/jplag) and [Dolos](https://dolos.ugent.be/) to analyze all submissions for an assignment.
+This pull request introduces the ability to pass custom arguments to these engines, allowing more control and flexibility in how similarity is calculated.
+
+As someone with no prior experience with either contributing to open source or in coding in the Go programming language, I was consistently encouraged and supported by my mentors, Lucas and Eriq, who gave me valuable guidance on writing cleaner and more efficient code. 
+This experience taught me about the importance of code quality and maintainability in production-level open-source collaborative projects.
+
+## Learning
+
+Over the past two months, my involvement in this project has been a period of immense growth and learning. 
+With the constant support and guidance of my mentors - Eriq Augustine and Lucas Ellenberger, I’ve had the chance to reflect on my skills, identify areas for improvement, and actively work on them.
+One of the biggest takeaways for me has been understanding the importance of writing clean, readable code - something I hadn’t fully appreciated before. 
+Their guidance has not only made me a better developer but have also shaped my growth more holistically. 
+I’ve started paying closer attention to deadlines, communicating more thoughtfully, and ensuring my work is both thorough and reliable.
+All in all, GSoC’25 has definitely proved to be a valuable experience for me.