mountain-ner-bert

bert-base-cased fine-tuned to tag mountain names in English text (BIO labels O, B-MOUNTAIN, I-MOUNTAIN). Code and data: https://github.com/Oleksandr190378/mountain-ner

Usage

from transformers import pipeline

ner = pipeline("token-classification", model="Oleks333/mountain-ner-bert", aggregation_strategy="first")
ner("Last summer we tried to climb Mont Blanc, but the weather near Matterhorn was terrible.")

Training data

~1,000 synthetic sentences with exactly one mountain each (names scraped from Wikipedia lists, sentences generated with an OpenAI model) plus ~200 sentences without mountains. Split 80/10/10 train/validation/test. 3 epochs, batch size 16, lr 3e-5.

Results

Entity-level (seqeval) on the held-out test split (120 sentences): precision 0.990, recall 1.000, F1 0.995.

Limitations

The test split comes from the same synthetic distribution, so real-world performance is lower. Observed failure modes: the second mountain in a sentence is sometimes missed, unseen multi-word names can be cut ("Fitz Roy" -> "Roy"), short names such as "K2" get low confidence.