03 · AI · Vision & OCR
Retail-Compliance VLM Fine-Tuning Program
A full applied-research loop on one 24 GB GPU: dataset design → multi-model fine-tuning → vLLM serving → quantitative and human evaluation → production deployment behind an unchanged API contract.
PeriodMar 2026StatusApplied research → productionAreaAI · Vision & OCR
Highlights
- Streaming, matrix-balanced downsampling of 99K→~40K examples with per-bucket caps; compact schema conversion.
- LoRA/4-bit fine-tunes of Qwen3-VL-4B, SmolVLM, Florence-2 and Qwen3.5-0.8B/2B; vision-token budgeting cut a run from ~115 h to ~28 h.
- vLLM serving (LoRA adapters → merged weights) with dynamic batching; production winner: Qwen3.5-2B.
- Flag-by-flag accuracy evaluation against Gemini on a held-out 10K-image set; a Tkinter GUI to audit results visually.