The forecast is created by a group of trained forecasters that collect and analyze a range of different data sources. Sources range from field observations submitted to the crowd sourcing platform regobs.no and metrological data provided by MET. In short, the daily forecast includes a short textual description of the major avalanche problem together with a number indicating the danger level (on a scale from 1-5).
The data sources used to determine the forecast include both numerical and text-based features. For example, the crowd sourced observations often include aspects of the snowpack, recorded according to a protocol, the result of snow instability tests, and a textual description of the situation in the mountains. As such, the task of creating a forecast can be framed as a generative task based on multiple data domains.
In this master thesis, the focus will be on exploring the possibility of generating automated forecasts. The training data will consist of openly available data from RegObs, metrological data, and historical forecasts. The scientific contribution will center on how these automated forecasts can be made trustworthy.
The general power of modern LLMs makes it likely that providing the abovementioned data as context is a more viable path in terms of raw performance than developing any model from scratch. An important part of this project is therefore to investigate how high-stake outputs, such as avalanche forecasts, should approach the inclusion of LLMs.
The specific study design is to be decided by the student. However, avalanche forecasters at NVE can be made accessible for evaluation or short study subject interviews.
Requirements:
- Proficiency in Norwegian.
- Interest in how LLMs should be integrated into high-stake applications, if at all.
- Proficiency in Python and knowledge about basic API functionality. Developing the proposed system will include incorporating multiple data sources.
Interest in more than “just” model development but also conducting a potential qualitative experiment with real forecasters as subjects.
If relevant, the student is encouraged to submit parts of their work as a research contribution to a fitting venue, either in machine learning or environmental hazards research.