README.md

metadata

license: mit
language:
  - fi
metrics:
  - f1
  - precision
  - recall
  - accuracy
base_model:
  - google-bert/bert-base-uncased
  - TurkuNLP/bert-base-finnish-cased-v1
pipeline_tag: text-classification
tags:
  - classification
  - news

News Relevancy Classifiers

FinBERT-ft-v3

Model Description

• Purpose: This model is trained for a specific task in research, it is not a commmercial product and should not be used in for-profit.
• Architecture: bert-base-finnish-cased-v1
• Fine-tuning task: Four-class Finnish news-headline relevancy classification
• Dataset: ~225 Finnish headlines (2024–2025) manually labeled into:
  • 0 — Not Relevant
  • 1 — Least Relevant
  • 2 — Highly Relevant
  • 3 — Most Relevant
• HF Repo: cloud0day3/finbert-ft-v3 (latest v4 checkpoint, 6 June 2025)
• Date Trained: 2025-06-06

Model Inputs

• A raw Finnish headline (string), truncated/padded to 96 tokens.
• Tokenization handled by the bundled vocab.txt + tokenizer_config.json + special_tokens_map.json.

Model Outputs

• A single integer label (0–3). Mapped to human-readable categories:

  LABELS = {
      0: "Not Relevant",
      1: "Least Relevant",
      2: "Highly Relevant",
      3: "Most Relevant"
  }
  

Intended Use

• Primary: Automatically assign a relevancy score to Finnish news headlines so that downstream pipelines (e.g., filtering, ranking) can operate without manual triage.

Examples of use:

• Pre-filtering a news aggregation feed.

• Prioritizing headlines for editorial review.

• Input to summarization/retrieval pipelines.

Out-of-Scope Uses

• Any non-Finnish text (e.g., English, Swedish).

• Multi-sentence inputs or full articles (this model is tuned on single-sentence headlines).

• Tasks other than relevancy (e.g., sentiment analysis, topic modeling).

• High-risk decision making without human oversight (e.g., emergency alerts).