Trailwise outdoor planner model configuration
Base model: Qwen/Qwen2.5-0.5B-Instruct, approximately 0.49B parameters. Unchanged open weights, Apache-2.0. No fine-tuning or adapter is claimed. This repository contains the application configuration and model card for the unchanged base-model option explicitly allowed by the hackathon rulebook. The weight files remain in the named upstream repository.
Load the base weights, not this configuration-only repository:
from transformers import AutoTokenizer, AutoModelForCausalLM
base = 'Qwen/Qwen2.5-0.5B-Instruct'
revision = '7ae557604adf67be50417f59c2c2f167def9a775'
tokenizer = AutoTokenizer.from_pretrained(base, revision=revision)
model = AutoModelForCausalLM.from_pretrained(base, revision=revision)
Before submission pin revision to the exact hash recorded in a successful
Trailwise response. Set the same hash in the Space's MODEL_REVISION variable.
The response reports actual loaded parameters and rejects models above 3B.
Application changes
No weight changes. Trailwise supplies intent extraction and candidate-selection
prompts, retrieves live map/weather/air-quality evidence, and checks activity
windows with deterministic constraints. LM Format Enforcer constrains JSON.
The SLM selects only an existing option from a checked shortlist within five
points of the highest planner score. Grounding rules check extraction against
the user's words. The final answer renders values from source records.
See application_config.json for runtime defaults.
Intended use and limits
Prototype outdoor planning for ten activities. It is not medical advice, a hazard-warning service, a traffic model, or an emergency navigation system. Extraction can fail, and small-model reasoning is limited. The application asks for clarification or withholds a recommendation when required data is missing. Opening hours, booking, route accessibility, local alerts, flash flooding and landslides require separate verification. Air quality is a model forecast, not a venue sensor observation. Downwind scoring is an uncalibrated heuristic.
Evaluation and disclosure
Unit tests cover map compatibility, wind direction, time windows, sensitive profiles and full-duration forecast checks. See the code repository's VALIDATION.md for actual local execution results. Synthetic tests do not establish field safety or performance on the judging queries. OpenAI Codex assisted with code and documentation. It is not called by the running assistant. Upstream model authors retain their original notices.
Application source and deployment
The complete application, tests, presentation and deployment instructions are in Trailwise-GeoAI-solution.zip. See PROJECT.md for application documentation and DEPLOY-MODAL.md for independent hosting. Application source is MIT licensed inside the archive and in CODE-LICENSE. The unchanged Qwen base model remains Apache-2.0. Public application: https://nimatullahkhan24--trailwise-geoai-web.us-east.modal.run
API: POST /ask on the same host. Cloud health, model readiness and clarification
behavior are verified. Full map queries can fail when public data providers time out.
Free hosting credits are finite and availability throughout judging is not guaranteed.