rpant/iolai26-solve

final submission

device map to gpu 0

Batch-1 no-padding generation; blend label-aware and positional answer alignment in lean

Lean mode: symbolic becomes fallback-only (model answers every item, fills blanks only)

Lean path: parse bare answer lines directly, skip marker/salvage and format induction (lean flag only)

Trim code comments to be concise and neutral

Switch LLM pass to the minimal-prompt (lean) path by default

Add toggle to disable match_letters assignment pass (default off) for isolation

Refine solver: greedy decoding config, lightweight linguistic-reasoning prompt, letter-matching via assignment, default device placement

Add 14B-AWQ shard 2

Add 14B-AWQ shard 1

Swap weights to Qwen2.5-14B-Instruct-AWQ (shard 3 + metadata)

Remove 7B weights ahead of 14B swap

Add SYMBOLIC_ONLY toggle; enable it for a symbolic-baseline diagnostic submission

Fix AWQ config warning: modules_to_not_convert null -> empty array

Fix output parsing that zeroed the score: tolerant markers, no prose fallback, salvage pass, adaptive modes

Fix sandbox OOM: token-budget batching, OOM retry/abstain, checkpointed submission writes

Ship Qwen2.5-7B-Instruct-AWQ weights at repo root (MODEL_ID='.')

Notebook-style MODEL_ID in script.py; move metrics into solver/; untrack dev-only eval/tests/data

Single-shot-with-tools v1: scaffolded prompts, shared pipeline, explanation track

Initial commit: IOL-AI 2026 solver