Verify Capacity-Driven Gains from Multilingual SFT at 27B

Larger models extract disproportionately more from multilingual supervised fine-tuning

4.52x
27B/4B Slope Ratio
3.42x
Relative Advantage 27B vs 4B
0.820
Quality 27B at 55 langs
p<0.001
Slope Test Significance

Scaling Curves: Quality vs Number of Languages

Language Group Slope Ratios (27B/4B)

Total Quality Gain by Model Size

Final Quality at 55 Languages

Language Group Analysis

GroupSlope (27B)Slope (4B)Ratio
High Resource0.004640.001124.15x
Mid Resource0.006320.001524.16x
Low Resource0.006240.001504.17x
Typologically Distant0.006260.001304.80x