ssb000ss/regional-sentiment-ru

Regional multilingual news sentiment

Fine-tuned XLM-RoBERTa-base for 3-class news headline sentiment in Russian, Ukrainian, Kazakh, Chinese, Japanese, Arabic, and Western European languages.

Why this model exists

The downstream OSINT dashboard scores ~200 news headlines every five minutes for its overall threat indicator. VADER (the upstream choice) returns 0.0 for every Cyrillic or CJK headline, which zeroed out 25 % of the threat formula for any non-English-dominant region. This fine-tune restores that signal for Cyrillic without regressing Latin text — English keeps VADER under the script-aware router in oracle_service.compute_sentiment.

Labels

id	name	what it covers
0	negative	casualty / conflict / disaster / strong-negative coverage
1	neutral	routine / procedural / non-evaluative coverage
2	positive	resolution / cooperation / strong-positive coverage

Training data

Mix of three public sources (~30.9 K train / 10.7 K val / 16.2 K test):

• MonoHime/ru_sentiment_dataset — 190 K RU reviews+news, label convention remapped 0/1/2 ↔ neg/neu/pos.
• tyqiangz/multilingual-sentiments — ZH (2.5 K), JA (2.5 K), AR/EN/ ES/DE/FR (~3.7 K each).
• cardiffnlp/tweet_sentiment_multilingual — Twitter sentiment in EN/FR/DE/ES/AR/HI/IT/PT.

Recipe: scripts/prepare_sentiment_dataset.py in the GitHub repo.

Training

• Apple-Silicon MPS, batch 8 + grad_accum 2 (effective 16), max_length 96, lr 1e-5, warmup 6 %, weight decay 0.01.
• HF Trainer, save_strategy="steps", save_steps = steps_per_epoch // 10 → checkpoint every 0.1 epoch.
• Early-stopped at 1.1 epoch (patience 4 from best @0.6 epoch).
• Recipe: scripts/train_sentiment.py in the GitHub repo.

Evaluation

• val_f1_macro (best checkpoint): 0.655
• test_f1_macro (final): 0.597
• test_accuracy (final): 0.600

Test is lower than val because the test split has a heavier Twitter/cardiff share where the Twitter-tuned baseline pre-train leaks through.

Smoke results vs VADER baseline

On a held-out hand-picked set of Russian news headlines spanning the three classes:

Class	VADER (baseline)	this model
Negative coverage	0.000 (VADER returns 0 for any Cyrillic input)	−0.45 to −0.75
Positive coverage	0.000	+0.40 to +0.55
Neutral / procedural coverage	0.000	≈ 0.00 (correctly close to zero)

Cyrillic gains are dramatic — the model recovers a signal that VADER silently dropped to zero. English and CJK have small regressions vs strong VADER and dictionary baselines on news-specific phrasings, so the regional-fork runtime routes only Cyrillic input through this model and keeps VADER + a hand-curated CJK character dict for the other scripts.

Usage in the regional fork

huggingface-cli login
uv run --group ml python scripts/download_model.py
# adapter auto-discovers backend/data/models/sentiment-finetuned/final/
SENTIMENT_ML_BACKEND=auto uv run uvicorn main:app --reload

Direct usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline
tok = AutoTokenizer.from_pretrained("ssb000ss/regional-sentiment-ru")
mdl = AutoModelForSequenceClassification.from_pretrained("ssb000ss/regional-sentiment-ru")
clf = pipeline("text-classification", model=mdl, tokenizer=tok)

# Pass a Russian / Ukrainian / Kazakh / Chinese / Japanese / Arabic /
# Spanish / German / French headline.  Output: 3-class sentiment with
# confidence in [0..1].
print(clf("<your headline here>"))
# [{'label': 'negative' | 'neutral' | 'positive', 'score': 0.5x}]

License

CC-BY-NC-4.0 (non-commercial), inherited from the regional fork it ships with.

Citation

If you use this model, please cite the upstream base model (cardiffnlp/twitter-xlm-roberta-base-sentiment-multilingual) and the public sentiment datasets listed under Training data above.