J-I-N: OCR and Cited RAG on AMD ROCm
Two Match solutions built on one model, Qwen3-VL-8B-Instruct (BF16, Hugging Face Transformers), on the mandated ROCm 10.0 base image, with the weights baked in so the container never needs the network.
Match 2 (OCR): reads US and Chinese licence plates, stop, speed-limit and work-zone signs from uncooperative images. A resident server loads the model once and warms up the GPU; each grader call to /app/app.py is a standard-library client that talks to it over a Unix socket and always writes the output JSON. The prompt and post-processing encode the transcription rules: drop US state names and slogans, keep the Chinese province character, join multi-line signs top to bottom. AMD GPU test: 10/10 official samples, 0.27 to 0.65 s per image, 17.7 GB peak VRAM.
Match 3 (RAG): answers questions over a mixed corpus (pdf, docx, xlsx, csv, logs, source code, images) with the value only and the exact set of source files. Every file is parsed defensively, images are transcribed by the same model during indexing, and retrieval is BM25 with identifiers kept as whole tokens. The model proposes the answer and its evidence; code then decides the citations: the value must literally appear in a cited file or the answer is refused, withdrawn documents are never used, and a second file is cited only when it supplies an identifier the question does not give. AMD GPU test: 10/10 sample questions with exact citations, about 1 s per question, 18 GB peak VRAM.
Code: github.com/awaisbari0408/jin-amd-challenge
.png&w=1920&q=75)
.png&w=1920&q=75)
.png&w=1920&q=75)
.png&w=1920&q=75)
