CampusClaim
Receipt-to-reimbursement for campus events: fine-tune Gemma 4 vision on SROIE, compare plain vs LoRA, Gradio claim draft.
CampusClaim is a receipt-to-reimbursement assistant for student clubs and campus events, submitted for Theme 3B (Receipt Information Extraction).
We fine-tune Google’s open-weight Gemma 4 E2B vision model with Unsloth LoRA on ICDAR 2019 SROIE receipts (company, date, address, total), then compare the plain and fine-tuned models on the same frozen 100-receipt test sample. Field-level metrics show gains especially on address and date. A Gradio app supports upload/camera → extract → edit → claim draft; adapters are saved for reload. Policy RAG over official campus finance PDFs is prepared as the Theme 3 combined-approach bonus (shortlisted documents and a 15-question citation/refusal test).
Repo default branch is dev. GPU training runs on Colab T4; local checks and the OCR fallback app stay CPU-only.
Built with
AI tools
Prize Categories Entered
- Submitted
Best Use of Gemma 4
We use Gemma 4 as a multimodal vision model: it reads receipt images and returns structured JSON fields. Fine-tuning with LoRA adapts that multimodal capability to campus reimbursement workflows, and the Gradio UI lets users correct fields before drafting a claim.
- Submitted
Best Open-Source AI Project
CampusClaim is built entirely on open-weight and open-source AI: Gemma 4, Unsloth LoRA, Hugging Face SROIE data, PEFT adapters, and a public GitHub repo with reproducible notebook, metrics, and Gradio demo—no closed API required for the core extraction path.
- Event prizeSubmitted
Theme 3: Fine-Tuning Gemma
CampusClaim fine-tunes Gemma 4 E2B with Unsloth LoRA on ICDAR 2019 SROIE receipt images for Theme 3B extraction (company, date, address, total). We compare plain vs fine-tuned models on the same frozen 100-receipt sample, ship a Gradio claim-draft app, and save adapters for reload—exactly the fine-tune, application, and comparative analysis the theme asks for.
