Vaarta SFT dataset v2: structured messages format, ~178K examples, multilingual

Vaarta SFT v5 (Gemma): k12×3 + Alpaca 30K + Wikipedia 15K. No Samanantar. Devanagari-filtered. Total 57,679.

Vaarta SFT v4: k12×3 (24,000) + Alpaca (30,000) + Wikipedia (16,367) + Samanantar (10,000). Total 80,367. Llama-3.2 chat template.

Vaarta Marathi SFT v3: k12-conversations (×3) + Marathi Wikipedia comprehension + Samanantar en→mr translation. 87,993 total examples. Native Llama-3.2 chat template.