Part of NewsIntel AI — a CPU-deployable, LLM-free pipeline that extracts structured public-safety events (accidents, disasters, crimes) from Bangladeshi news in Bengali and English. This model is stage 3 · target-conditioned sentence selection of that chain:

document → relevance gate → event-type router → evidence selection
         → NER → relation extraction → knowledge graph → structured event JSON

Model details


Base model	`xlm-roberta-base`
Task	`evidence`
Input	`text-pair`
Max length	192
Format	ONNX INT8 (dynamic quantization), ~279 MB
Version	`c749dd7cf2726`
Languages	Bengali (primary, ~96% of training corpus), English

Labels

not_evidence
evidence

Thresholds

None — this model emits raw scores; the caller ranks or thresholds.

Decision rule

score = softmax(logits)[1] over (condition, sentence); rank top-k

These values also ship machine-readable in model_manifest.json, so a serving process can consume the model without hardcoding anything.

Evaluation

Metric	Value
ROC-AUC	0.8942
PR-AUC	0.782
F1	0.6865

Usage

python

from huggingface_hub import snapshot_download
import onnxruntime as ort, numpy as np
from transformers import AutoTokenizer

d = snapshot_download("saidylive/newsintel-evidence", revision="c749dd7cf2726",
                      allow_patterns=["model_int8.onnx", "*.json", "*.model"])
tok = AutoTokenizer.from_pretrained(d)
sess = ort.InferenceSession(f"{d}/model_int8.onnx", providers=["CPUExecutionProvider"])

enc = tok("type: ATTRIBUTE | target: death_count | need:  | event: accident",
            "সাভারে সংঘর্ষে ৩ জন নিহত হয়েছেন।", return_tensors="np")
logits = sess.run(None, {k: v for k, v in enc.items()
                         if k in {i.name for i in sess.get_inputs()}})[0]
# decision rule (from model_manifest.json):
#   score = softmax(logits)[1] over (condition, sentence); rank top-k

Training data & provenance

Trained on the bd_eng_news_daily Kaggle corpus of Bangladeshi news (~713k articles, ~96% Bengali by character ratio). Labels are silver, not human-annotated: a teacher LLM produced structured event annotations, which were distilled into these small models. No manual annotation was performed at any stage.

This matters for how you read the metrics: they measure agreement with LLM-generated labels, not with human ground truth. There is no human-labelled evaluation set.

Limitations & bias

Silver labels cap the ceiling. Systematic teacher-LLM errors are inherited.
Domain-specific. Tuned to Bangladeshi public-safety news; expect degradation on other domains, regions, or registers.
Opinion pieces leak through. Editorials and foreign wire stories are sometimes classified as events by the upstream gate/router.
Entity noise. NER tags some generic Bengali nouns (e.g. রাজধানীর "of the capital", সদর "HQ") as locations.
No calibration. Confidence-style outputs are uncalibrated; do not read them as probabilities of correctness.
INT8 quantization trades a little accuracy for ~4× size reduction and CPU speed.
Not for high-stakes use. Casualty counts and event classifications are unverified model output and must not be used for emergency response, journalism, or policy without human review.

License

Released under cc-by-nc-4.0 — free to share and adapt for non-commercial purposes with attribution. Note that the training corpus consists of copyrighted news articles and the labels were LLM-distilled; downstream users are responsible for their own compliance.

Citation

bibtex

@software{newsintel_ai,
  title  = {NewsIntel AI: distilled multilingual event extraction for Bangladeshi news},
  author = {Md. Sheikh Saidy},
  year   = {2026},
  url    = {https://huggingface.co/saidylive/newsintel-evidence}
}