jeonghoonlee/iol-ai-2026-clean-entry

IOL-AI 2026 entry

Submission code for the IOL-AI Challenge 2026.

How it runs

1. script.py reads /tmp/data/test.csv.
2. The solver uses deterministic row rules where they return a complete result.
3. Remaining rows use the packaged Qwen model. For translations into English, rows whose numbered query items contain explicit hyphenation use a draft-and-audit generation path. Both generations use the base model instance loaded by solve_rows. Other translation rows use the single-item path.
4. A final repair checks complete translation cells against visible row evidence. It first copies an explicitly displayed right-hand English phrase only when the query line is shaped like (n) | English phrase. It then runs the repeated-correspondence repair from the previous candidate. It does not read reference answers or another model candidate.
5. The repaired result is written to submission.csv with columns id, pred, explanation.

Model files

Path	Role
models/Qwen2.5-14B-Instruct-AWQ	primary solve stage

The neural matching pass is disabled. No second language model or learned selector is included. The base model folder includes its upstream LICENSE (Apache-2.0).

Output format

• pred: JSON list of strings, one answer per numbered item, in order.
• explanation: short plain note for the optional jury track (not part of the automatic score).

Runtime

The packaged model loads once from local files. The repair stage is CPU-only and runs after Qwen exits. The Qwen-only 10-row practice run completed in 231 seconds on a Titan RTX with a 10,462 MiB peak. T4 runtime has not been measured.