Hitelcy/sarvix-multilingual-1

Sarvix Multilingual 1

Fine-tuned from Qwen2.5-1.5B-Instruct using LoRA (r=16, alpha=32) on 768 assert-only examples across 13 languages. The model takes a vague, ambiguous, or emotionally loaded message and confidently restates what the sender most likely meant — without asking any questions. It always responds in the same language as the input.

Behavior

• Input: a short, vague, or emotionally loaded chat message
• Output: a confident restatement of the likely intended meaning
• Never asks questions — always commits to an interpretation
• Responds 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: 768 assert-only examples across 13 languages
• Languages: English, Spanish, French, German, Portuguese, Japanese, Arabic, Hindi, Swahili, Italian, Russian, Korean, Chinese (Mandarin)

Known limitations

• Swahili, Hindi, and Korean may produce less fluent output due to lower training coverage compared to the other languages.