DialSeg-Ar - LDC Telephone: leaderboard
Linear topic segmentation of conversational Arabic from the DialSeg-Ar benchmark (Chirkunov et al., ACL 2026): the model splits an ordered list of utterances into topic-coherent segments. Test split of Gulf, Iraqi and Levantine telephone conversation transcripts. WindowDiff on a 0-1 scale, the share of sliding windows where the predicted boundary count differs from the gold count; lower is better.
Metric: WindowDiff (0-1, lower is better; share of sliding windows over the utterance sequence whose predicted boundary count differs from the gold count; test split of Gulf, Iraqi and Levantine telephone conversation transcripts; LLMs prompted zero-shot, DialSeg-Ar-Gemma3-4B fine-tuned on the training split). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 3 4B (IT) | 0.52 | #3848 |
| 2 | Fanar-1-9B-Instruct | 0.67 | #1765 |
| 3 | ALLaM-7B-Instruct-preview | 0.77 | #1861 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dialseg-ar-ldc-telephone · How It Works · Data refreshed daily, snapshot 2026-11-12.