Spaces:

---

Keeby-smilyai

/

LLM-kitchen

like 2

Sleeping

GPU + Storage Grant Request for Educational Space — LLM Kitchen 🍳

Hi Hugging Face team 👋

I’m BC, a student and open‑source contributor passionate about making AI education more transparent, hands‑on, and joyful. My mission is simple: AI for everyone — and for everyone to truly understand how LLMs work, from the inside out.

I’m writing to request a physical GPU grant (A10G, T4, or L4) along with persistent storage to support my personal project, LLM Kitchen.

LLM Kitchen is a curriculum‑rich Space that helps beginners train and publish small language models from scratch. It’s built around playful metaphors—like Auto‑Seasoning™ for hyperparameter tuning and Inference Kitchen for testing outputs—to make model training feel intuitive and creative. The interface guides users through architecture selection, hyperparameter configuration, backend training, inference testing, and optional publishing to the Hub. It’s designed to be friendly for students, educators, and curious newcomers alike.

This is a personal project by me (BC), not affiliated with SmilyAI. While I’m incredibly grateful for the ZeroGPU grant that powers another Space I maintain, ZeroGPU is too limited for this kind of educational workflow.

I understand Hugging Face doesn’t typically grant GPUs for training, but I want to clarify that LLM Kitchen is not a production trainer. It’s an educational sandbox designed to help learners understand how language models work.All training is ephemeral unless published, capped in size, and protected by a 48-hour timeout. The goal isn’t optimization—it’s transparency, reproducibility, and curriculum-style learning. I’ve built in safeguards and playful scaffolding to make this a responsible, beginner-friendly teaching tool.

Why ZeroGPU Falls Short for LLM Kitchen

• 48‑hour timeout: To conserve resources, I’ve implemented a 48‑hour training limit, but ZeroGPU’s shorter execution windows still interrupt curriculum‑style workflows.
• No wasted Hub resources: The Space stores nothing on the Hub unless the user explicitly chooses to publish. Without persistent storage, all training is ephemeral.
• Model size constraints: Even with models capped at 4–8 layers and small batch sizes, backend execution is essential for curriculum ramping, norm tracking, and diagnostic probes—none of which are feasible under ZeroGPU’s constraints.

🎓 Why This Grant Matters

With a physical GPU and storage, I could:

• Run persistent backend training with curriculum‑style logging, norm evolution tracking, and latent reasoning probes
• Let users publish trained models directly to the Hugging Face Hub with custom model cards
• Expand to support word problems, multi‑step reasoning, and CSV trace logging for early/mid/late inference chains
• Avoid accidental quota drain and timeouts, making the Space stable for classrooms and community demos
• Build reproducible workflows that show how architecture, hyperparameters, and training dynamics affect model behavior

🌍 Impact

LLM Kitchen is not just a tool—it’s a gateway to understanding AI. With Hugging Face’s support, it could:

• Reach hundreds of students in classrooms and online workshops within the first year
• Serve as a hands‑on lab for educators introducing AI concepts in an accessible, playful way
• Empower self‑learners worldwide to go beyond “using” AI and instead understand its inner workings
• Foster a community of curious builders who share their trained models, insights, and experiments openly

By making the inner workings of LLMs visible and approachable, we can empower every learner, in every corner of the world, to not just use AI — but to understand it deeply.

⚙️ Technical Readiness

LLM Kitchen is already fully scaffolded and operational in a limited form:

• UI & UX: Complete, with guided steps for architecture selection, hyperparameter tuning, backend training, inference testing, and publishing
• Backend logic: Implemented with curriculum ramping, norm tracking, and diagnostic probes ready to run on a physical GPU
• Publishing flow: Integrated with Hugging Face Hub for optional model uploads and custom model card generation
• Resource safeguards: Model size caps, batch size limits, and a 48‑hour timeout to ensure responsible usage
• Educational content: Playful metaphors, inline explanations, and visual feedback designed for beginners and classrooms

With hardware in place, I can immediately enable persistent training, richer diagnostics, and reproducible workflows—no major development delays required.

Thank you for considering this request—and for all the encouragement and tools you provide to the open‑source community. I’m happy to share walkthroughs, logs, or demo links if helpful.

Warmly,
BC