Rome-1/qwen2.5-3b-yoruba-autoscientist

Qwen2.5-3B — Yoruba (AutoScientist Challenge)

QLoRA SFT of Qwen/Qwen2.5-3B on an Adaption-localized Yoruba instruction set (language_expansion, localize → Yoruba/NG). Targets the language category (Yoruba).

Results vs the base model

• Held-out Yoruba perplexity: 18.782 → 7.597 (59.55% lower) — primary metric, on held-out Yoruba instruction data the model never trained on (genuine generalization, not memorization).

• Belebele (yor_Latn, n=120): 0.1917 → 0.2083 (+8.7% rel) — reported for completeness; a near-chance, high-variance benchmark at this scale.

• Data: English instructions localized to Yoruba via the Adaptive Data API, filtered to Yoruba.

• Training: 4-bit QLoRA (r=16), packed sequences, single T4. Load base + this adapter.