27 lines
1.1 KiB
Markdown
27 lines
1.1 KiB
Markdown
|
---
|
||
|
|
base_model: Qwen/Qwen2.5-14B-Instruct-AWQ
|
||
|
|
license: apache-2.0
|
||
|
|
license_link: https://huggingface.co/Qwen/Qwen2.5-14B-Instruct-AWQ/blob/main/LICENSE
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
tags:
|
||
|
|
- competition
|
||
|
|
- iol-ai-2026
|
||
|
|
---
|
||
|
|
|
||
|
|
# IOL-AI 2026 Challenge submission
|
||
|
|
|
||
|
|
This repo is a redistribution of [Qwen/Qwen2.5-14B-Instruct-AWQ](https://huggingface.co/Qwen/Qwen2.5-14B-Instruct-AWQ)
|
||
|
|
(Apache 2.0 license) plus a `script.py` for the
|
||
|
|
[IOL-AI 2026 Challenge](https://huggingface.co/spaces/iol-ai-challenge/iol-ai-2026).
|
||
|
|
|
||
|
|
## Approach
|
||
|
|
|
||
|
|
`script.py` prompts the model per-problem with a structured "rule table" method (enumerate every
|
||
|
|
example, derive a rule table, answer only from that table), followed by a self-review pass where the
|
||
|
|
model checks its own draft against the rule table and the original examples before finalizing. It
|
||
|
|
tracks a wall-clock budget against the competition's 30-minute limit and falls back to a single fast
|
||
|
|
greedy pass (or, if time is fully exhausted, a placeholder answer) rather than risk the process being
|
||
|
|
killed with no `submission.csv` written.
|
||
|
|
|
||
|
|
No training or fine-tuning is involved — this is inference-only, as required by the competition rules.
|