metadata
license: cc-by-nc-4.0
language:
- is
pipeline_tag: text-classification
library_name: transformers
tags:
- icelandic
- sentiment-analysis
- text-classification
- sequence-classification
- social-media
sources: >-
Risamálheildin slices of forums/blogs, manually labelled by us, and our own
small corpus made from samples gathered from social media
Task: 3-class sentiment analysis → ["negative", "neutral", "positive"]
Base model: mideind/IceBERT-igc (Icelandic RoBERTa)
TL;DR
A small Icelandic RoBERTa fine-tuned for 3-way sentiment on non-ironic text. Pairs well after an irony gate (first run the irony model; only classify sentiment if not_ironic).
How to use
from transformers import AutoTokenizer, AutoModelForSequenceClassification
model_id = "ambj24/icelandic-sentiment"
tok = AutoTokenizer.from_pretrained(model_id)
mod = AutoModelForSequenceClassification.from_pretrained(model_id)
text = "Þjónustan var frábær!"
inputs = tok(text, return_tensors="pt")
probs = mod(**inputs).logits.softmax(-1).tolist()[0]
labels = ["negative", "neutral", "positive"]
print(dict(zip(labels, probs)))
Input length: short posts; trained with max length ~128 tokens.
Data: social-media style Icelandic.
Domain shift: trained on short, informal posts.
Positive/neutral/negative labels; only examples judged not ironic.
Typical setup: 3 epochs, LR ≈ 2e-5, batch ≈ 16, max length 128.