---
license: apache-2.0
base_model: Qwen/Qwen2.5-7B-Instruct
tags:
  - iol-ai-2026
---

# IOL-AI 2026 submission (v3)

Submission for the [IOL-AI 2026 challenge](https://iolai.org/):
Qwen2.5-7B-Instruct quantized to 4-bit (bitsandbytes NF4) at load time,
greedy decoding, with per-item prompting.

v3: 7B in 4-bit instead of 1.5B (fp16 fallback if bitsandbytes is missing).

v2 changes over the plain baseline:

- one generation per numbered item instead of one per problem
- brief step-by-step reasoning, answer extracted from a `FINAL:` line
- task-type-specific answer-format instructions
- a one-shot worked example in the prompt
- a global time budget that shrinks reasoning space near the 30-min limit

- `script.py` reads `/tmp/data/test.csv` and writes `submission.csv`
  (columns `id`, `pred`, where `pred` is a JSON list of per-item answers).
- Model weights are shipped in this repo and loaded from `"."` because the
  eval sandbox has no internet.

Base model: [Qwen/Qwen2.5-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct)
(Apache 2.0).