luxia-21.4B-alignment-v1.2: benchmark results
Saltlux's 21.4B DPO-aligned model on a passthrough depth-upscale of InternLM2-20B; its v1.x GSM8K gains drew contamination flags on the Open LLM Leaderboard (2024). Provider: Saltlux. Released 2024-10-25.
Aggregate Rating 1448 ± 17, rank #1012 of 1656 rated models, from 35 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard v1 - HellaSwag | 90.86 | Normalized accuracy (%) (10-shot) | 99.8 |
| Open Chinese LLM - WinoGrande | 73.72 | Accuracy (%) | 99.7 |
| Open LLM Leaderboard v1 - ARC Challenge | 77.73 | Normalized accuracy (%) (25-shot) | 99.7 |
| Open LLM Leaderboard v1 - TruthfulQA MC2 | 79.16 | MC2 (%) (0-shot) | 99.6 |
| Open Chinese LLM - TruthfulQA MC | 69.57 | Accuracy (%) | 99.4 |
| Open LLM Leaderboard v1 - WinoGrande | 86.27 | Accuracy (%) (5-shot) | 99.1 |
| Open Chinese LLM - HellaSwag | 77.12 | Accuracy (%) | 98.8 |
| Open Chinese LLM - ARC Challenge | 64.25 | Accuracy (%) | 95.9 |
| Open Chinese LLM Leaderboard | 69.42 | Official Nonzero-Task Mean (%) | 93.8 |
| Open LLM Leaderboard v1 - MMLU | 67.86 | Accuracy (%) (5-shot) | 90 |
| Open LLM Leaderboard - BBH | 63.71 | Score | 87.9 |
| Open Korean LLM - KorNAT-Harmless | 66.68 | Normalized accuracy (%) | 86.6 |
Interactive version: theaggregate.ai/model?slug=luxia-21-4b-alignment-v1-2 · How It Works · Data refreshed daily, snapshot 2026-10-20.