Hacktoberfest Hack Day Tumakur x upGrad School of Technology

Build an original project that uses open-source or open-weight AI as an important part of how it works. Teams could create an agent skill, build with an open-weight large or small language model, or build or adapt an open-source model harness.

It’s time to see how much you can build with a lightweight, open model. Gemma 4 packs multimodal intelligence into an open-weights model that you can access through the Gemini API. So, what can Gemma 4 bring to your project?

• Multimodal Experience: Work with text and images to create an assistant that understands more of what users share.
• Focused AI Tools: Build a focused AI tool for learning, creativity, productivity, or your community.
• Rapid Prototyping: Experiment with an open model while using the Gemini API to move quickly from an idea to a working prototype.

Bring your idea to life… what will you build with Gemma 4 today?

1. Background
   Open-weight AI models are models whose weights are publicly released, so anyone can download, run, study and build on them. Organisations such as Google, OpenAI and Alibaba release open-weight models under licences such as Apache 2.0.

Choosing a model for a real application means understanding both what the model is (its size, design, licence and stated limitations) and how well it performs on the tasks that matter. Published benchmarks give one view. Careful, hands-on testing by the people who will use the model gives another. This process is called model evaluation.

2. Objective
   Research and evaluate a set of open-weight models, using the free web playgrounds and the models’ published documentation, and present evidence-based findings and recommendations.

1. Background
   Language models have broad general knowledge but no access to private or specialised documents, such as institutional regulations, course material or project documentation. When asked about such documents, a model may produce answers that sound convincing but are false.

Retrieval-Augmented Generation (RAG) addresses this. The application first retrieves the passages of a document most relevant to a question, then gives only those passages to the model and instructs it to answer from them, with citations.

2. Objective
   Using the starter kit, build a RAG application with Google’s open-weight Gemma 4 model for one of the two problem statements below. The application must answer accurately from the provided documents, cite its sources, and decline questions the documents do not cover.

1. Background
   A general-purpose model can be adapted to a specific task by fine-tuning: continuing its training on examples of that task. Fine-tuning is useful when a task needs a consistent format, specialised terminology or behaviour that prompting alone does not reliably produce.

LoRA (Low-Rank Adaptation) makes fine-tuning practical on limited hardware. It leaves the original model weights unchanged and trains a small set of additional weights. With Unsloth, Google’s open-weight Gemma 4 model can be fine-tuned on a free Colab GPU.

A fine-tuned model is only useful if it is measurably better than the original. Every fine-tuning project therefore needs a fair comparison against the plain model.

2. Objective
   Fine-tune Gemma 4 for one of the two problem statements below, build an application around the fine-tuned model, and present a comparative analysis of the plain and fine-tuned models.