Modelpublic

rpant/iolai26-solve

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes69downloads

Model Card

IOL-AI 2026 Solver

Solves IOL-style linguistics puzzles (Linguini CSV) with a single-shot LLM. Ships Qwen/Qwen2.5-14B-Instruct-AWQ (4-bit AWQ, Apache-2.0) at the repo root, loaded from MODEL_ID = "."; runs on a T4 within the 30-minute budget.

Architecture

—`script.py` — entrypoint: reads /tmp/data/test.csv, writes submission.csv (id, pred JSON list, explanation). Config flags at the top; submission written before the model loads and after every step.
—`solver/pipeline.py` — orchestration. The model answers every puzzle from a minimal prompt (no scaffold, no chain-of-thought); output is parsed into one answer per item and aligned by position. A light greedy-anchored self-consistency vote refines answers while the clock allows. A deterministic symbolic layer (solver/) is a last-resort fallback only.
—`solver/llm.py` — batched transformers/AWQ generation, greedy, repetition_penalty=1.0, per-token deadline.
—`solver/direct.py` — prompt and answer parsing.

Run

bash

python3 script.py [test.csv] [submission.csv]   # defaults: /tmp/data/test.csv, submission.csv

Weights are the unmodified Qwen/Qwen2.5-14B-Instruct-AWQ release (Apache-2.0; LICENSE retained).