In this work, we tackle some major challenges in Arabic multi-label emotion classification especially the issues of class imbalance and label correlation that often hurt model performance, particularly for minority emotions.
Our approach:
Stacked contextual embeddings from fine-tuned ArabicBERT, MarBERT, and AraBERT models.
A meta-learning strategy that builds richer representations.
A hybrid loss function combining class weighting, label correlation matrices, and contrastive learning to better handle class imbalances.
π Extensive experiments show significant improvements across Precision, Recall, F1-Score, Jaccard Accuracy, and Hamming Loss. π The hybrid loss function in particular helped close the gap between majority and minority classes!
We also performed ablation studies to break down each componentβs contribution and the results consistently validated our design choices.
This framework isn't just for Arabic it offers a generalizable path for improving multi-label emotion classification in other low-resource languages and domains.
Big thanks to my co-authors: Muhammad Azeem Aslam, Wang Jun, Nisar Ahmed, Li Yanan, Hu Hongfei, Wang Shiyu, and Xin Liu!
Would love to hear your thoughts on this work! π
π Takeaways 1οΈβ£ Even the best computer-using models struggle with simple, context-dependent tasks.Β 2οΈβ£ Visual perception and reasoning remain major hurdles for multimodal agents. 3οΈβ£ Real-world use cases reveal significant gaps between hype and reality. Perception accuracy drops to near zero by the last turn π
We just crossed 1,500,000 public models on Hugging Face (and 500k spaces, 330k datasets, 50k papers). One new repository is created every 15 seconds. Congratulations all!
3 replies
Β·
reacted to BrigitteTousi's
post with πabout 2 months ago
Another impressive model that joined the ranking today is ALLaM-AI/ALLaM-7B-Instruct-preview. After a long wait finally ALLaM is here and it is IMPRESSIVE given its size !
Google just released PaliGemma 2 Mix: new versatile instruction vision language models π₯
> Three new models: 3B, 10B, 28B with res 224, 448 π > Can do vision language tasks with open-ended prompts, understand documents, and segment or detect anything π€―
π Excited to share our technical report on the Southeast Asian multilingual model Sailor2 and its latest updates!
Our 49-page report details Sailor2's development journey, including multilingual data cleaning, small model data mixture simulations, multi-stage continual pre-training, multi-stage post-training, and multi-cultural multi-lingual evaluations. Sailor2 aims to streamline the multilingual model pre-training process efficiently for the community.
π§ We highlight Sailor2's impressive performance in low-resource language translation scenarios and its cultural understanding advantages in Southeast Asia, promoting practical applications for regional languages.
Model updates include:Β π‘ More precise outputs: Reduced redundancy in model outputs through refined post-training data and optimization techniques.Β π Handling longer texts: Expanded to handle up to 128K context length in Southeast Asian languages through long-text training.Β β‘οΈ Faster inference: Achieved 2.5x faster inference speed with speculative decoding.Β πͺοΈ More model sizes: Introduced new sizes of 3B and 14B through model pruning.
π All models are Apache-licensed for commercial use; development tools (code, resources) are open-source.
π HuggingFace Spaces Ranking Tracker - Your Complete AI Trend Analytics!
Introducing the Spaces Ranking Tracker, a comprehensive analytics dashboard that tracks and analyzes every AI application in the HuggingFace ecosystem.
β¨ Key Features: β’ Real-time tracking of daily ranking changes over 30 days β’ Detailed analysis of top 100 trending spaces β’ User-based integrated score visualization β’ One-click access to space details β’ Interactive rank change graphs
π Dashboard Components: 1. Main Dashboard - Daily rank trend graphs - Top 20 creators' combined score chart - Detailed space information cards - Real-time trending score updates
2. Space Detailed Analysis - Creation date, current rank, and trending score - 30-day ranking history - Direct space access - Custom color coding for intuitive rank display
π― How to Use: β’ Monitor latest AI community trends β’ Track your project's performance β’ Discover popular AI demos β’ Analyze competing projects β’ Follow AI ecosystem dynamics
3. Interactive Features - Custom filtering options - Sorting by various metrics - Detailed performance statistics - Comprehensive trending scores - Historical data tracking
Stay on top of every movement in the HuggingFace ecosystem with daily ranking updates! π Try it now!
There's so much you could do with these developments. Especially combining them together into agentic applications or fine-tuning them on your use case.
1 reply
Β·
reacted to AdinaY's
post with π₯π§ 3 months ago
DeepSeek-R1 & DeepSeek-R1-Zero: two 660B reasoning models are here, alongside 6 distilled dense models (based on Llama & Qwen) for the community! deepseek-ai deepseek-ai/DeepSeek-R1
β¨ MIT License : enabling distillation for custom models β¨ 32B & 70B models match OpenAI o1-mini in multiple capabilities β¨ API live now! Access Chain of Thought reasoning with model='deepseek-reasoner'
reacted to MohamedRashad's
post with β€οΈ4 months ago
3C3H AraGen Leaderboard welcomes today deepseek-ai/DeepSeek-V3 and 12 other models (including the late gpt-3.5 π) to the ranking of best LLMs in Arabic !
Observations: - DeepSeek-v3 ranked 3rd and only Open model among the top 5 !
- A 14B open model (Qwen/Qwen2.5-14B-Instruct) outperforms gpt-3.5-turbo-0125 (from last year). This shows how much we came in advancing and supporting Arabic presence within the LLM ecosystem !
- Contrary to what observed in likelihood-acc leaderboards (like OALL/Open-Arabic-LLM-Leaderboard) further finetuned models like maldv/Qwentile2.5-32B-Instruct actually decreased the performance compared to the original model Qwen/Qwen2.5-32B-Instruct. It's worth to note that the decrease is statiscally insignificant which imply that at best, the out-domain finetuning do not really hurts the model original capabilities acquired during pretraining. Previous work addressed this (finetuning VS pretraining) but more investigation in this regard is required (any PhDs here ? This could be your question ...)
π―Fine-tuning SmolLM2 on a lightweight synthetic reasoning dataset for reasoning-specific tasks. Future updates will focus on lightweight, blazing-fast reasoning models. Until then, check out the blog for fine-tuning details.
