named places
Specific mountains, corries, glens and passes appear across the record.
Cairngorm Club journals · 1893–2023
Building a text-dataset of Scotland’s mountain snow
A human-verified record of snow and ice observations recovered from historic mountaineering journals—made searchable by place, season and year.
2,381
curated observations
1618–2018
dated record span
16
mapped regions
96.3%
with map coordinates
Cairngorms Western Massif · photograph by Craig Aitchison
Text is an under-utilised asset for Scottish climate change
Two centuries of observations, mined using AI and verified by human volunteers.
Mountaineers recorded snowfields, cornices, avalanches, frozen lochs in journals and diaries well before the deliberate action of the late 20th and 21st centuries. SNOSCOT brings those dispersed accounts together as structured observations while retaining the original language of the source material. We used AI to mine the text (wince), but hand verified all extracted fields against the reference text. This project would not have been possible without AI.
The collection centres on Cairngorm Club journals published from 1893 to 2023 and includes additional historical references reaching back to 1618. It is a record of what writers noticed and what survives in the archive—not a continuous instrument record.
Interactive atlas
Find the observation behind the point.
Search the full text, choose a region, or narrow the dated record. Map points represent broad regional groupings rather than exact observation sites.
Loading observations…
Exploratory data analysis
What is inside the archive?
A descriptive view of when observations appear, where they were assigned, and the language used to characterise snow and ice.
mean snow score
The curation score describes how positive or substantial the snow description is. The "Snow Score" is itself defined and calculated by GPT4o.
years represented
Coverage is highly uneven and follows the availability and content of source journals.
most-recorded month
465 observations with an April date—the strongest monthly concentration.
01 · Temporal coverage
Observations by decade
The high early-20th-century count reflects journal coverage as well as what was observed.
02 · Seasonality
Observations by month
Monthly analysis uses 1,997 valid month values.
03 · Geography
Most represented regions
04 · Curation score
Score distribution
A high score signals a strongly positive or substantial snow description, not data quality.
05 · Vocabulary
Frequent extracted terms
06 · Data completeness
Useful context, with visible gaps.
The original quotation and regional assignment are almost complete. Exact dates and point-level geography are less consistent in historical prose, so missingness is retained rather than guessed.
From page to observation
Machine-assisted discovery.
Human-verified evidence.
The project was designed to make a very large historical archive tractable without treating model output as ground truth. Language models identified candidate passages; people checked, corrected and contextualised every retained observation.
-
01
Acquire & extract
Journal PDFs are collected and their pages converted into machine-readable text. This involves a process known as document-chunking where overlapping chunks are generated for downstream extraction. This is because LLMs have a context window
-
02
Identify candidates
A GPT-4o-family model finds compact passages that explicitly mention snow, ice or related cryosphere conditions. Why GPT-4o? Fundamentally, cost. SimonF92 funded API calls with his own money. Price estimates to redo this with advanced models are ~£150-£250.
-
03
Verify & repair
Each quotation, entity, score, date and location is reviewed against the source text by a human being; hallucinations are removed, dates are refined and locations are narrowed down as much as possible.
-
04
Map & publish
Specific place names are reconciled into broad geographic groups and released with the curated output.
Dataset structure
Twelve fields keep the quote and its context together.
The quotation remains the central unit. Derived labels make the archive searchable, while comments preserve decisions that need qualification.
| Field | What it contains |
|---|---|
text | The retained verbatim passage from the source. |
entity | Extracted snow and ice terms, sometimes comma-separated. |
score | A 0–10 description score, from absent/melted to substantial or strongly positive snow. |
date | The best available original date string, including partial dates. |
annotator_comment | Human notes about uncertainty, derivation or historical context. |
general_location | A consistent broad mountain region used by the atlas. |
specific_location | The named mountain, corrie, loch, glen or pass. |
year · season · month · day | Parsed date components where the source supports them. |
Coordinates | A representative coordinate for the broad mapped region. |
Open research dataset
Collaboration is key
The mined and curated dataset outputs are released under Apache 2.0. Rights in the original journal PDFs remain with the Cairngorm Club; contact the relevant rights holder before reusing source documents themselves. If an academic institution wishes to adopt and refine this work with larger models for a PhD student project then please do get in touch.
Dataset DOI
10.57967/hf/7619
Ben Macdui · photograph by Peter Hudson