Week 14 | Presentations

This week was really exciting because we had two project presentations lined up — one from our classmates on Preswald, and one from our team on Huggingface. —

Preswald Presentation

Our classmates introduced Preswald, a lightweight framework for building data-visualization dashboards using Python and HTML. Rather than open-sourcing every feature, Preswald follows an Open Core model: the core library is free and extensible, but advanced widgets and enterprise integrations require a paid license.

Key takeaways from their demo:

• Community health
  • A small but active subreddit keeps conversation going.
  • The official Slack channel exists, but it’s under-moderated and often slow to respond.
• Core contributions
  1. LaTeX support – rendering math in charts
  2. Sidebar functionality – handling tokenized menu items proved tricky
  3. Interactive chat interface – built against their published docs
  4. Debug panel – streamlined development and visual testing

They were candid about the hurdles they encountered:

• Sparse documentation made onboarding the codebase a steep climb.
• Rapid churn in upstream releases frequently broke their local branches.
• Community support was uneven, leaving some PRs in limbo.

Tip: Schedule quick sync-calls with Preswald maintainers and advocate for a formal Contributing Guide to smooth the initial setup process.

---

Huggingface Presentation

Our team’s presentation covered our final OSSD project on Huggingface 🤗 — one of the largest open-source AI ecosystems. We structured our talk into five parts:

1. What Is Huggingface?
Huggingface provides a unified hub for ML models, datasets, and tools.

• Transformers (e.g., DeepSeek, Llama, Nemotron)
• Diffusers (e.g., Stable Diffusion XL, ControlNet)
• Datasets (e.g., Wikipedia, CommonsenseQA, COCO)

2. Contribution Statistics

• Total contributions: 22
  • 11 PRs (4 merged)
  • 11 issues
• Breakdown by category:
  • Feature Implementation — 3
  • Bug Fixes — 3
  • Bug Analysis — 4
  • Documentation Improvement — 4
  • Community Support — 3
  • Infrastructure Improvement — 3
  • Test Additions — 2

3. Our Workflow

• Slack coordination: Clear channels for feature requests, reviews, and hotfixes
• Branch strategy: Each contributor claimed an issue, created a feature branch, then pinged reviewers
• Pull-request etiquette: Descriptive titles, linked issues, and inline code comments kept reviews efficient

Individual Contributions

• Haocheng Lu: Patched indexing logic, optimized attention masks, corrected padding in the image processor, fixed documentation links, and added configurable head dimensions for MoE models.
• Yufeng Xu: Closed outdated bug branches, enforced type consistency, stabilized KV-cache parameters, and improved the installation guide.
• Haochen Yang: Refactored complex code paths for clarity and reliability, and patched edge-case behaviors in the training loop.
• Minjun Zhu: Enhanced documentation clarity, wrote comprehensive tests for cache updates and tensor-parallel features, added “resume checkpoint” support in ClearML, and provided example code for fine-tuning SigLIP2.

Group Contribution: Time-Based Strategy

• Motivation: Step/epoch-based evaluation and saving can drift across different hardware environments.
• Process:
  1. Claimed the feature request for eval_strategy=TIME.
  2. Iterated with maintainers on PR formatting and unit tests.
  3. Verified through three test scenarios:
     • Evaluation trigger (eval_minutes=1)
     • Checkpoint saving (save_minutes=1)
     • Logging output (logging_minutes=1)

---

Final Thoughts

Comparing the two projects highlighted the spectrum of open-source engagement:

• Preswald is nimble and still carving out its ecosystem, but needs stronger documentation and community processes.
• Huggingface boasts a mature infrastructure with clear contribution pathways, yet still relies on dedicated testers and communicators.

No matter the scale, impactful open-source work comes down to:

1. Deep codebase familiarity
2. Transparent communication
3. Rigorous testing
4. Persistence through setbacks

Really proud of everyone’s efforts this week! 🎉