Outdoor Window model configuration
This repository describes unchanged Qwen2.5-1.5B-Instruct used by Outdoor Window.
No training or measured fine-tuning improvement is claimed. Approximately 1.54B parameters.
Runtime used by the application: Ollama qwen2.5:1.5b, Q4_K_M.
Load the exact base model
from transformers import AutoTokenizer, AutoModelForCausalLM
import json
cfg=json.load(open('base_model.json'))
tokenizer=AutoTokenizer.from_pretrained(cfg['base_model'], revision=cfg['revision'])
model=AutoModelForCausalLM.from_pretrained(cfg['base_model'], revision=cfg['revision'])
The repository contains usage configuration rather than redistributed weights.
This follows the rulebook's unchanged-base-model provision. The base loads through
from_pretrained; this configuration repository by itself is not a standalone weights repository.
If judges require a standalone loadable repository, upload a licensed copy of the base weights and tokenizer.
Use
The SLM extracts a request into JSON and chooses among live-data candidate IDs and evidence IDs. Python validates the result and produces factual text from those selected evidence records. The model has no direct access to arbitrary URLs or tools. User queries and mapped names are untrusted data.
Limitations
Extraction can fail. Forecasts and mapped data can be incomplete. No medical clearance, live traffic, verified opening hours or confirmed booking availability is provided. Conservative prototype screening rules are not clinical thresholds. India-only prototype; Asia/Kolkata.
Disclosure
Developed with OpenAI Codex as a coding assistant. No OpenAI model is called by the running app. Data: OpenStreetMap contributors, Open-Meteo and CAMS. Optional routing: OpenRouteService. The third-party weekend-planner repository was reviewed for ideas; no code was copied.