---
license: apache-2.0
library_name: transformers
base_model: google/gemma-4-12B
pipeline_tag: text-generation
tags:
  - gemma4
  - iol-ai-2026
---

# IOL-AI 2026 — gemma4:12b hybrid (symbolic numeral solver + budget-managed LLM)

Three-pass pipeline:

1. **Symbolic pass** — `numeral_solver.py` brute-forces numeral systems
   (base, word values, word-order convention) from the problem's own
   examples and answers `text_to_num` / `num_to_text` exactly when a
   consistent system exists; otherwise falls through to the LLM.
2. **Baseline pass** — a fast low-token answer for every remaining
   problem, with `submission.csv` atomically checkpointed after every row.
3. **Upgrade pass** — per-row time-sliced reasoning with per-task-type
   method prompts; a baseline answer is replaced only when the upgrade
   parses to the correct number of answers. Translation reasoning is
   deliberately capped low: on gemma4-12B, truncated reasoning plus a
   forced short answer scored measurably higher than completed reasoning.

Model: gemma4-12B (Apache 2.0), weights shipped in this repo, loaded 4-bit
via bitsandbytes at startup (fp16 12B does not fit the 16 GB T4).

## Upload checklist (before submitting)

1. Put the gemma4-12B weight files (safetensors + config + tokenizer) in
   the repo root so `script.py` loads them from `"."`.
2. Repo must be **public** at submission time.
3. Submit via the competition Space; the platform mounts the hidden test
   set at `/tmp/data/test.csv` and runs `script.py`.