Swahili math models no longer chase only the answer: Fikiria gates its reward on the reasoning trace, and its 14B model hits 79.2% on AfriMGSM
Synopsis
Fikiria changes what Swahili mathematical reasoning is rewarded for: instead of scoring only the final answer, its reward multiplies a format gate by a weighted sum of exact-match accuracy and the ChrF++ similarity of the Swahili reasoning to an in-language reference chain. The two-stage recipe runs cold-start supervised fine-tuning before GRPO, and Fikiria-AfriqueQwen-14B reaches 79.2% on the AfriMGSM Swahili test set, against 68.0% for GPT-4o under the same prompt and strict scorer; in a small human evaluation, however, GPT-4o's outputs were preferred overall.
Interpretation
For Swahili math problems, the training target moves from rewarding only a correct final answer to also rewarding a reasoning chain written in Swahili and close in surface form to a reference chain. Earlier multilingual math-reasoning methods optimize answer accuracy, or enforce language consistency only on the final answer, so a correct answer reached through English reasoning can still earn full reward. On the AfriMGSM Swahili test set, Fikiria-AfriqueQwen-14B reaches 79.2%, while GPT-4o reaches 68.0% under the same fixed prompt and strict answer-format scorer, partly because it often wraps answers in extra characters.
Training proceeds in two steps: supervised fine-tuning first teaches output structure and Swahili instruction following, then GRPO optimizes the reasoning trace under the gated reward. The reward is a format gate multiplied by a weighted sum of two terms: exact-match answer accuracy and the ChrF++ similarity of the reasoning to a Swahili reference chain, where ChrF++ is a parameter-free character n-gram overlap metric; the gate requires tags, an equation in the rationale and sufficient length, and zeroes the reward otherwise. Ablations on Qwen3 show supervised fine-tuning plus GRPO beating GRPO alone by 6.4 points at 8B and 12.0 points at 14B, but only 2.0 points at 32B, within noise; each configuration was a single training run.
Human evaluation reverses the automatic ranking: GPT-4o's reasoning outputs were preferred overall, while Fikiria was chosen more often for the naturalness of its Swahili. Automatic scoring measures answer correctness and surface language overlap, not whether individual reasoning steps actually support the answer, so a human comparison is needed for that dimension. Fifty-four samples were rated blind and pairwise by two native-Swahili-speaking co-authors, with 41% inter-annotator agreement and very different use of the tie label (51% versus 0%), so the direction of preference is robust but its magnitude is only approximate.
Perspective
The demonstrated scope is Swahili grade-school math, evaluated on the human-translated AfriMGSM test set of 250 examples; training data comes from translating GSM8K with GPT-4.1 and filtering with SSA-COMET-MTL at a 0.65 threshold, yielding about 2.3k training and about 400 test entries. A reader interested in other African languages or other low-resource reasoning tasks should treat these results as evidence for one language and one task.
ChrF++ measures surface overlap with a reference and does not verify that individual reasoning steps support the answer; the reward design itself was not ablated, so component contributions are not isolated; all trained results come from a single run per configuration, and no GRPO-only ablation was run on AfriqueQwen; AfriMGSM derives from GSM8K, which may be present in base-model pre-training data, so some accuracy may reflect contamination; training data is machine-translated and may carry systematic artifacts and mistranslations near the threshold; the human evaluation covers only 54 samples, is scored by co-authors and has 41% agreement, so its conclusions are suggestive.
