This week, I watched the presentations from 6 different groups in class. Also, I continued my contributions to Huggingface and personal contribution projects. Click to see the details!

Reflection on Presentations

This week, 6 groups presented their brilliant contributions to a variety of different open source projects. The one that gave me the deepest impression was Lucide, an open source website where you can find a myriad of icon resources. The two things that impressed me the most were (1) the astounding number of stars of the project (18.9K), and (2) how rude the maintainers were😂. I still remember one of the maintainers said “I won’t bother to count how many rules were violated in this PR”, which is - understandable, if their pursuit is to build a website where all contributors are highly professional and the styles of the icons are very consistent. However, in that case the project is more like a cathedral, not a bazaar of coffeehouse. In that case, the appearance that the project is welcoming to contributions is a disguise, and including the covenant code of conduct as a part of the codebase is sarcastic.
Another project I found really interesting is the Autogen. The goal of project is to give users an easy interface to build multi AI agent systems, and I like it that they focused on both frontend and backend contributions. One question I’m pretty curious about is how compatible the project is overall, because currently there are tons of different frameworks for AI models/agents. Autogen also reminds me of another project called aisuite, which aims to create a unified framework for all different kinds of models.

Open-source Contribution Progress

At this point, I have 3 PRs in progress, and I’ve been trying to find time to solve each of them:

• lm-evaluation-harness PR #2892: lm-evaluation-harness is a unified evaluation framework that is compatible with a huge number of different models and datasets, which saves the pain of writing evaluation code by yourself. However, currently the project only supoorts a limited number of datasets, and I worked to add AIME - a commonly used dataset for evaluating math and reasoning abilities of models - to the project. However, after submitting the PR I noticed this is more complicated than I expected.
• foundation-model-stack PR #390 I submitted a PR to fix a docstring in the codebase, which is an easy fix. However, there ia a DCO error in my PR, and I’m still figuring out how to fix it.
• huggingface transformers PR #37476 This is a PR aiming to fix incorrect installation guide in huggingface transformers, which I believe will be of great impact. However, after submitting the PR I was criticized by the user who opened the issue because I violated the DRY principle for software development, introduced a lot of redundancy in my PR and would cause trouble for future maintenance. Still working to figure out what will be the best solution.
• huggingface transformers PR #37642 We’ve addressed all the concerns of the reviewer, but haven’t received further responses from the community.