🚀 Official Release: AetherMind-KD-Student (184M)

I’m proud to release AetherMind-KD-Student, a compact 184M NLI model distilled from a DeBERTa-v3 teacher.
It achieves strong results across core, adversarial, and zero-shot NLI benchmarks.

🔍 Highlights

• 308 samples/sec on RTX 3050
• MNLI ~90.5%, SNLI ~89%
• Adversarial NLI:
  • ANLI R1: 73.6%
  • ANLI R2: 57.7%
  • ANLI R3: 53.67%
• Zero-shot:
  • SciTail: 78.83%
  • RTE: ~86.3%
  • HANS: ~77.7%
  • XNLI (EN): 90.92%

🧠 Training Data

SNLI, MNLI, ANLI (R1–R3)

🧪 Zero-Shot Generalization

RTE, HANS, SciTail, XNLI (English)

🙏 Credits

Developed by Sameer S. Najm
AetherMind project – Sam IT Solutions, Iraq

Open for collaboration and feedback!