Synthetic Curriculum & SnapPO for Low-Resource Languages

Evaluating the Solar methodology across 10 languages spanning five resource tiers, from English (5000T tokens) to Dzongkha (0.02T tokens).

+20.6

High-Resource Gain

+16.9

Mid-High Gain

+10.0

Low-Resource Gain

+10.0

Very-Low Gain

0.406

Min Transfer Ratio

Performance by Resource Tier

Full Pipeline Gain by Resource Tier

Component Contribution by Tier

Full Pipeline NLU Scores by Language

Transfer Ratio (vs Korean) - General NLU

Detailed Results

Full Pipeline Scores by Language

Language	Tier	Corpus (T)	NLU	Gen Quality	Reasoning
English	High	5000	99.4	97.1	94.4
Korean	Mid-High	320	82.3	77.5	76.4
Turkish	Mid	85	75.1	70.7	68.0
Vietnamese	Mid	78	72.4	70.4	65.1
Swahili	Low	4.5	53.7	50.9	47.4
Yoruba	Low	1.2	45.1	42.6	39.4
Quechua	Very-Low	0.15	35.2	32.0	30.1
Guarani	Very-Low	0.08	30.6	28.7	26.2
Bambara	Very-Low	0.04	26.1	25.8	21.1
Dzongkha	Very-Low	0.02	22.7	21.7	17.6

Gains by Resource Tier

Tier	SynCurr	SnapPO	Full Pipeline
High	+14.16	+9.39	+20.61
Mid-High	+11.60	+6.99	+16.94
Mid	+10.08	+5.90	+15.30
Low	+6.79	+5.06	+10.04
Very-Low	+7.11	+4.66	+10.01