QCRI/LlamaLens-Arabic

LlamaLens: Specialized Multilingual LLM Dataset

Overview

LlamaLens is a specialized multilingual LLM designed for analyzing news and social media content. It focuses on 19 NLP tasks, leveraging 52 datasets across Arabic, English, and Hindi.

LlamaLens

This repo includes scripts needed to run our full pipeline, including data preprocessing and sampling, instruction dataset creation, model fine-tuning, inference and evaluation.

Features

• Multilingual support (Arabic, English, Hindi)
• 19 NLP tasks with 52 datasets
• Optimized for news and social media content analysis

📂 Dataset Overview

Arabic Datasets

Task	Dataset	# Labels	# Train	# Test	# Dev
Attentionworthiness	CT22Attentionworthy	9	2,470	1,186	1,071
Checkworthiness	CT24_T1	2	22,403	500	1,093
Claim	CT22Claim	2	3,513	1,248	339
Cyberbullying	ArCyc_CB	2	3,145	900	451
Emotion	Emotional-Tone	8	7,024	2,009	1,005
Emotion	NewsHeadline	7	939	323	160
Factuality	Arafacts	5	4,354	1,245	623
Factuality	COVID19Factuality	2	3,513	988	339
Harmful	CT22Harmful	2	2,484	1,201	1,076
Hate Speech	annotated-hatetweets-4-classes	4	210,526	100,565	90,544
Hate Speech	OSACT4SubtaskB	2	4,778	1,827	2,048
News Genre Categorization	ASND	10	74,496	21,942	11,136
News Genre Categorization	SANADAkhbarona	7	62,210	7,824	7,824
News Genre Categorization	SANADAlArabiya	6	56,967	7,123	7,120
News Genre Categorization	SANADAlkhaleej	7	36,391	4,550	4,550
News Genre Categorization	UltimateDataset	10	133,036	38,456	19,269
News Credibility	NewsCredibilityDataset	2	8,671	2,730	1,426
Summarization	xlsum	--	37,425	4,689	4,689
Offensive Language	ArCyc_OFF	2	3,138	900	450
Offensive Language	OSACT4SubtaskA	2	4,780	1,827	2,047
Propaganda	ArPro	2	6,002	1,326	672
Sarcasm	ArSarcasm-v2	2	8,749	2,996	3,761
Sentiment	ar_reviews_100k	3	69,998	20,000	10,000
Sentiment	ArSAS	4	13,883	3,976	1,987
Stance	Mawqif-Arabic-Stance-main	2	3,162	560	950
Stance	stance	3	2,652	379	755
Subjectivity	ThatiAR	2	2,446	748	467

Results

Below, we present the performance of LlamaLens in Arabic compared to existing SOTA (if available) and the Llama-Instruct baseline, The “Delta” column here is calculated as (LLamalens – SOTA).

---

Task	Dataset	Metric	SOTA	Llama-instruct	LLamalens	Delta (LLamalens - SOTA)
News Summarization	xlsum	R-2	0.137	0.034	0.075	-0.062
News Genre	ASND	Ma-F1	0.770	0.587	0.938	0.168
News Genre	SANADAkhbarona	Acc	0.940	0.784	0.922	-0.018
News Genre	SANADAlArabiya	Acc	0.974	0.893	0.986	0.012
News Genre	SANADAlkhaleej	Acc	0.986	0.865	0.967	-0.019
News Genre	UltimateDataset	Ma-F1	0.970	0.376	0.883	-0.087
News Credibility	NewsCredibility	Acc	0.899	0.455	0.494	-0.405
Emotion	Emotional-Tone	W-F1	0.658	0.358	0.748	0.090
Emotion	NewsHeadline	Acc	1.000	0.406	0.551	-0.449
Sarcasm	ArSarcasm-v2	F1_Pos	0.584	0.477	0.307	-0.277
Sentiment	ar_reviews_100k	F1_Pos	–	0.343	0.665	–
Sentiment	ArSAS	Acc	0.920	0.603	0.795	-0.125
Stance	stance	Ma-F1	0.767	0.608	0.936	0.169
Stance	Mawqif-Arabic-Stance	Ma-F1	0.789	0.764	0.867	0.078
Att.worthiness	CT22Attentionworthy	W-F1	0.412	0.158	0.544	0.132
Checkworthiness	CT24_T1	F1_Pos	0.569	0.404	0.877	0.308
Claim	CT22Claim	Acc	0.703	0.581	0.778	0.075
Factuality	Arafacts	Mi-F1	0.850	0.210	0.534	-0.316
Factuality	COVID19Factuality	W-F1	0.831	0.492	0.781	-0.050
Propaganda	ArPro	Mi-F1	0.767	0.597	0.762	-0.005
Cyberbullying	ArCyc_CB	Acc	0.863	0.766	0.753	-0.110
Harmfulness	CT22Harmful	F1_Pos	0.557	0.507	0.508	-0.049
Hate Speech	annotated-hatetweets-4	W-F1	0.630	0.257	0.549	-0.081
Hate Speech	OSACT4SubtaskB	Mi-F1	0.950	0.819	0.802	-0.148
Offensive	ArCyc_OFF	Ma-F1	0.878	0.489	0.652	-0.226
Offensive	OSACT4SubtaskA	Ma-F1	0.905	0.782	0.899	-0.006

File Format

Each JSONL file in the dataset follows a structured format with the following fields:

• id: Unique identifier for each data entry.
• original_id: Identifier from the original dataset, if available.
• input: The original text that needs to be analyzed.
• output: The label assigned to the text after analysis.
• dataset: Name of the dataset the entry belongs.
• task: The specific task type.
• lang: The language of the input text.
• instructions: A brief set of instructions describing how the text should be labeled.
• text: A formatted structure including instructions and response for the task in a conversation format between the system, user, and assistant, showing the decision process.

Example entry in JSONL file:

{
    "id": "d1662e29-11cf-45cb-bf89-fa5cd993bc78",
    "original_id": "nan",
    "input": "الدفاع الجوي السوري يتصدى لهجوم صاروخي على قاعدة جوية في حمص",
    "output": "not_claim",
    "dataset": "ans-claim",
    "task": "Claim detection",
    "lang": "ar",
    "instructions": "Analyze the given text and label it as 'claim' if it includes a factual statement that can be verified, or 'not_claim' if it's not a checkable assertion. Return only the label without any explanation, justification or additional text.",
    "text": "<|begin_of_text|><|start_header_id|>system<|end_header_id|>You are a social media expert providing accurate analysis and insights.<|eot_id|><|start_header_id|>user<|end_header_id|>Analyze the given text and label it as 'claim' if it includes a factual statement that can be verified, or 'not_claim' if it's not a checkable assertion. Return only the label without any explanation, justification or additional text.\ninput: الدفاع الجوي السوري يتصدى لهجوم صاروخي على قاعدة جوية في حمص\nlabel: <|eot_id|><|start_header_id|>assistant<|end_header_id|>not_claim<|eot_id|><|end_of_text|>"
}

Model

LlamaLens on Hugging Face

Replication Scripts

LlamaLens GitHub Repository

📢 Citation

If you use this dataset, please cite our paper:

@article{kmainasi2024llamalensspecializedmultilingualllm,
  title={LlamaLens: Specialized Multilingual LLM for Analyzing News and Social Media Content},
  author={Mohamed Bayan Kmainasi and Ali Ezzat Shahroor and Maram Hasanain and Sahinur Rahman Laskar and Naeemul Hassan and Firoj Alam},
  year={2024},
  journal={arXiv preprint arXiv:2410.15308},
  volume={},
  number={},
  pages={},
  url={https://arxiv.org/abs/2410.15308},
  eprint={2410.15308},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}