fawoenix/silent-witness

Silent Witness

A Multimodal Multilingual Dataset for Human Rights Violation Detection

Dataset Summary

Silent Witness is a passage-level multi-label classification dataset for human rights violation detection in formal institutional documents. Each instance pairs an extracted text passage with the rendered page image from which it originates, annotated against an 11-category violation taxonomy derived from OHCHR international human rights law.

Source material is drawn from United Nations bodies (OHCHR) and established NGOs (Human Rights Watch, Amnesty International, FIDH).

Languages

Code	Language	Script	Source
ar	Arabic (MSA)	Arabic	OHCHR
fr	French	Latin	FIDH, Amnesty International
en	English	Latin	Human Rights Watch, Amnesty

Dataset Structure

Data Fields

Field	Type	Description
passage_id	string	Unique passage identifier encoding doc, page, index
doc_id	string	Source document identifier
page_number	int	Source page number (1-indexed)
language	string	ISO 639-1 / ISO 639-3 language code
text	string	Extracted passage text (min 30 words, max 2000 chars)
page_image	string	Path to rendered page PNG (150 DPI)
labels	list[string]	Multi-hot violation category labels
confidence	string	Annotation confidence: high / medium / low
source_org	string	Source organization
geographic_scope	string	ISO 3166-1 alpha-2 country code
synthetic	bool	true for synthetic instances only

Data Splits

Split	Passages	Notes
Train	1,243	Stratified by language
Validation	155	Stratified by language
Test	158	Stratified by language
Total	1,556

Violation Taxonomy

11-category label set derived from OHCHR top-level categories. Labels are not mutually exclusive — each passage receives a multi-hot vector. ENVIRONMENTAL_RIGHTS added based on corpus analysis (present in 40% of documents, grounded in UN Resolution 76/300).

Label	Description
TORTURE_ILL_TREATMENT	Torture, cruel, inhuman, or degrading treatment
ARBITRARY_DETENTION	Detention without legal basis or due process
EXTRAJUDICIAL_KILLING	Killing outside any legal framework
ENFORCED_DISAPPEARANCE	Abduction by state agents with denial of custody
FREEDOM_OF_EXPRESSION	Suppression of speech, press, or information
FREEDOM_OF_ASSEMBLY	Suppression of peaceful assembly or association
DISPLACEMENT_FORCED	Forced displacement, expulsion, or relocation
SEXUAL_GENDER_BASED_VIOLENCE	Sexual violence or gender-based persecution
DISCRIMINATION	Discriminatory treatment based on protected characteristics
FAIR_TRIAL_VIOLATION	Denial of fair trial or judicial guarantees
ENVIRONMENTAL_RIGHTS	Violations of the right to a clean, healthy, and sustainable environment

Label Distribution (full corpus)

Label	Count
DISCRIMINATION	1,307
FAIR_TRIAL_VIOLATION	949
ARBITRARY_DETENTION	676
TORTURE_ILL_TREATMENT	674
DISPLACEMENT_FORCED	576
SEXUAL_GENDER_BASED_VIOLENCE	530
EXTRAJUDICIAL_KILLING	497
ENFORCED_DISAPPEARANCE	397
FREEDOM_OF_ASSEMBLY	381
FREEDOM_OF_EXPRESSION	329
ENVIRONMENTAL_RIGHTS	59

Annotation

Annotator: Single primary annotator with multilingual background (Arabic, French, English).

Inter-annotator agreement (15% sample, n=233 passages, second annotator):

Label	Cohen's Kappa	Interpretation
TORTURE_ILL_TREATMENT	0.778	Substantial
ARBITRARY_DETENTION	0.799	Substantial
EXTRAJUDICIAL_KILLING	0.867	Substantial
ENFORCED_DISAPPEARANCE	0.812	Substantial
FREEDOM_OF_EXPRESSION	0.847	Substantial
FREEDOM_OF_ASSEMBLY	0.814	Substantial
DISPLACEMENT_FORCED	0.789	Substantial
SEXUAL_GENDER_BASED_VIOLENCE	0.908	Substantial
DISCRIMINATION	0.524	Moderate
FAIR_TRIAL_VIOLATION	0.697	Substantial
ENVIRONMENTAL_RIGHTS	0.675	Substantial
Overall (macro avg)	0.774	Substantial

DISCRIMINATION scores moderate agreement due to conceptual overlap with multiple other categories. This is documented as a known limitation.

Baseline Results

Three baseline experiments on the test set:

Experiment	Model	Test F1 Macro	Test F1 Micro
E1 — Text-only	XLM-RoBERTa-base	0.7931	0.8799
E2 — Multimodal	XLM-RoBERTa-base + ViT-base (frozen)	0.8070	0.8942
E3 — Zero-shot	CLIP ViT-H-14 (no fine-tuning)	0.5123	0.5474

Per-Language Test F1 Macro

Language	E1 XLM-R	E2 XLM-R + ViT	E3 CLIP zero-shot
Arabic (MSA)	0.5854	0.5962	0.5649
English	0.6659	0.6659	0.5631
French	0.4877	0.5197	0.3067

Key findings:

• Multimodal model (E2) consistently outperforms text-only (E1) across all languages. Visual page features provide signal beyond text alone.
• French F1 is consistently the lowest across all experiments. Likely cause: FIDH legal French terminology is underrepresented relative to English and Arabic in XLM-R pretraining.
• Zero-shot CLIP (E3) achieves high recall but low precision at optimal threshold (0.1), indicating the model labels most passages positively. Fine-tuning is necessary for reliable detection.
• ENVIRONMENTAL_RIGHTS scores near 0.00 in E1 and E2 due to very low test support (7 passages). Addressed in v1.0 with expanded corpus.

Sources

UN Bodies

Organization	Portal	Languages
OHCHR	ohchr.org/en/documents	AR

NGOs

Organization	Portal	Languages
Human Rights Watch	hrw.org/reports	EN
Amnesty International	amnesty.org/en/documents	FR
FIDH	fidh.org/en/issues	FR

Known Limitations

• Single primary annotator for the majority of the corpus. Inter-annotator agreement measured on 15% sample only.
• Page-level image pairing: one visual unit may span multiple unrelated topics. Individual image extraction deferred to v1.0.
• ENVIRONMENTAL_RIGHTS has very low support (7 test passages) — insufficient for reliable model learning at pilot scale.
• No negative examples (passages with no violation content) included. Balanced negative class scoped to v1.0.
• French performance gap (F1 Macro ~0.49–0.52) warrants investigation in v1.0 with expanded French corpus.

Versioning

Version	Scope
v0.1 (current)	Pilot: 1,556 passages, 3 languages, 11 labels, 3 baselines
v1.0 (planned)	Expanded corpus, negative class, second full annotator, Darija extension
v2.0 (planned)	Individual image extraction, additional MENA languages, OCR for scanned Arabic

Citation

@dataset{silent_witness_2026,
  author    = {BEN SALEM, Lamia},
  title     = {Silent Witness: A Multimodal Multilingual Dataset for Human Rights Violation Detection},
  year      = {2026},
  publisher = {HuggingFace Datasets},
  url       = {https://huggingface.co/datasets/fawoenix/silent-witness}
}

License

Source documents are used under the following policies:

• UN documents: public domain (ST/AI/189/Add.9/Rev.2)
• Human Rights Watch: open access
• Amnesty International: open access
• FIDH: open access