A small Arabic reasoning model: Qwen2.5-3B-Instruct fine-tuned with LoRA to reason step-by-step in Arabic inside native … tags before answering. Trained end-to-end on a Mac with MLX, built to run on a laptop.

Results 15-question Arabic eval (eval/eval_set.jsonl, greedy decoding, substring match on the final answer), 2026-09-05:

Model Score Qwen2.5-3B-Instruct (base) ¹¹⁄₁₅ = 73.3% Wamda-3B (LoRA, 600 iters) ¹³⁄₁₅ = 86.7% Gains vs base (+4): the 3-pill trap (1 hour, not 1.5), the boxes word problem (18 SAR), age algebra (Sara = 21), discount+VAT (3680). The structure does its job on multi-step problems the base model fumbles.

Regressions vs base (−2): raw multi-digit multiplication the base got right — 47 × 36 and 13 × 17. A common first-pass SFT trade-off: the format-tuning improves reasoning structure while narrow synthetic data can cost raw arithmetic. Fix: more pure-arithmetic drills (see Roadmap).

Fore more - https://github.com/seyhunak/wamda-3b-reasoning and https://huggingface.co/seyhunak/Wamda-3B-Reasoning