Jidayi/ruhu-operator-adtc-2026

ADTC 2026 Laptop LLM Challenge — Qwen3-4B-IQ4_XS: offline business operations for inventory-led African SMEs # Ruhu Operator — ADTC 2026 Laptop LLM Challenge **Team:** `ruhu-operator` · **Domain:** `corporate_enterprise` · **Model:** `Qwen3-4B-IQ4_XS` An offline business operations assistant for inventory-led African SMEs — FMCG wholesalers, mini-markets, hardware dealers, and automotive-parts dealers whose records already live in a POS export, a spreadsheet, or a small database. The model reads those records from the prompt and returns grounded analysis, correct arithmetic, and structured output. It runs entirely on the owner's laptop, with no network dependency at any point after the weights are fetched. See REPORT.md for the technical writeup. ## Quick start ```bash bash download_model.sh ``` Downloads `Qwen3-4B-IQ4_XS.gguf` (2.1 GB) into `model/` and verifies it against a pinned SHA-256. The script is idempotent and needs no credentials. Validate with the official profiler: ```bash adtc-profiler run --submission . --mode participant \ --output submission.json --skip-accuracy ``` ## Measured results Official `adtc-profiler` run on the target profile — Intel Xeon at 2.80 GHz, 4 physical cores with SMT disabled, 7.8 GB RAM, no GPU, Ubuntu 22.04.5 LTS: | measurement | value | |---|---:| | generation throughput | 9.37 tokens/s | | peak RSS | 2,396 MB | | `arc_easy` `acc_norm` | 0.72 (50 samples) | | thermal throttling | none observed | | parameter count | 4,022,468,096 — matches the `4B` claim | ## Design summary Three decisions shaped the submission, each measured rather than argued: - **Qwen3-4B over Qwen3-1.7B.** A 26.7-point accuracy gap on a purpose-built 20-prompt evaluation set. The gap is substantive: the 1.7B miscomputes VAT-exclusive amounts, classifies a supplier payment as a staff cost, and names a SKU with 75 weeks of cover as an urgent stockout risk. Reduced throughput was accepted to protect the 50%-weighted accuracy component. - **IQ4_XS over Q4_K_M.** Identical answers on all 20 evaluation prompts, but 2.32 GB peak resident memory against 4.1 …