Hitelcy/sarvix-multilingual-1

Sarvix Multilingual 1

Fine-tuned from Qwen2.5-1.5B-Instruct using LoRA (r=16, alpha=32) on 804 examples across 12 languages. Trains the model to either ask a clarifying question on ambiguous emotional input, or assert a confident reflective interpretation when the input contains clear emotional content — and to respond in the same language as the input.

Training

• Base: Qwen/Qwen2.5-1.5B-Instruct
• Method: LoRA (target modules: q_proj, k_proj, v_proj, o_proj)
• Epochs: 3
• Dataset: 804 examples (347 English + 457 multilingual across 11 additional languages)
• Final training loss: 1.33
• Mean token accuracy: 0.79

Known limitations

• Swahili, Hindi, and Korean have fewer training examples (15-25 each) and may produce less fluent output compared to Spanish, French, German, Portuguese, Italian, Russian, Arabic, and Japanese, which have stronger coverage.