Glonce

ZeroR system takes 2nd place in Nepali meme hate speech challenge

researchThe Glonce editorsarXiv🔥50Glonce rating50Impact45Novelty55Relevance65Surprise25Players15Credibility85

A team describes its entry to the CHiPSAL 2026 shared task on multimodal hate speech and sentiment detection in Nepali memes, which covered two subtasks: binary hate speech classification and three-class sentiment analysis. The system, named ZeroR, adapts the RA-HMD (Robust Adaptation of Hateful Meme Detection) framework around Qwen3-VL-8B-Instruct, a vision-language model with native support for the Devanagari script used to write Nepali. Because the model reads Devanagari text directly from meme images, the authors say their pipeline avoids the error propagation that separate OCR and translation steps would otherwise introduce. Training runs in two stages: first, LoRA fine-tuning with an added MLP projection head, framing classification as a generative task; second, contrastive fine-tuning of the backbone with a supervised InfoNCE loss. To cope with class imbalance in the data, the team applied minority-class oversampling, image augmentation and focal loss. At inference time, the system combines the token probabilities from the Stage 1 model with the classifier scores from Stage 2, weighted by values tuned on a validation set. The submission scored an F1 of 0.797 on hate speech detection, good for 2nd place in the shared task, and an F1 of 0.518 on sentiment analysis, good for 4th place. The paper also reports ablation studies and error analysis aimed at informing future work on adapting large vision-language models to low-resource South Asian languages, though it does not name the authors, their institution, the exact competition date, or the size and source of the Nepali meme dataset used.

Key facts

  • Built for the CHiPSAL 2026 shared task on multimodal hate speech and sentiment detection in Nepali memes, covering binary hate speech classification and three-class sentiment analysis
  • Adapts the RA-HMD framework to Qwen3-VL-8B-Instruct, a vision-language model with native Devanagari support, avoiding separate OCR and translation steps
  • Uses a two-stage pipeline: LoRA fine-tuning with an MLP projection head, then contrastive backbone fine-tuning with a supervised InfoNCE loss
  • Handles class imbalance with minority oversampling, image augmentation and focal loss, and ensembles Stage 1 and Stage 2 scores at inference with validation-tuned weights
  • Achieved F1 0.797 for 2nd place on hate speech detection and F1 0.518 for 4th place on sentiment analysis

Why it matters

Most hateful-meme detection work targets English and other high-resource languages. This entry tackles Nepali, a low-resource South Asian language written in Devanagari, and does so by reading the script directly out of a vision-language model rather than bolting on separate OCR and translation stages that tend to compound errors.

Who it affects

The direct audience is researchers and shared-task participants working on multimodal content moderation for low-resource languages, plus the CHiPSAL 2026 organizers who run the hate speech and sentiment benchmarks this system was scored against. Platforms that need to moderate Nepali-language content are the indirect beneficiaries of any technique that generalizes.

How to use it

The described recipe is a two-stage adaptation of Qwen3-VL-8B-Instruct: LoRA fine-tuning with an MLP projection head for generative classification, followed by contrastive fine-tuning of the backbone with a supervised InfoNCE loss, then an inference-time ensemble of the two stages' scores using validation-tuned weights. Class imbalance is addressed with minority oversampling, image augmentation and focal loss. The paper is a competition writeup rather than a released product; no code, model weights or pricing are mentioned in the text.

How solid is it

The results are concrete and competition-scored: an F1 of 0.797 placed the system 2nd on hate speech detection, and an F1 of 0.518 placed it 4th on sentiment analysis, both measured against other CHiPSAL 2026 entrants. The authors say they include ablations and error analysis, though the source text does not spell out the ablation results, the dataset size, or how the other teams' scores compared.

Risks and caveats

The source gives no author names or institutional affiliation, no exact date for the shared task or submission, and no detail on the size, source or composition of the Nepali meme dataset. It also does not explain what the original RA-HMD framework looked like before this adaptation, and it does not name or score the competing teams beyond the paper's own rankings, so the margin against 1st place and the teams ahead and behind it are unknown.