parkky21/orpheus-3b-hi-ft-1e

parkky21/orpheus-3b-hi-duo-voices (अनुष्का • करन)

🔎 Model Summary

• Base model: canopylabs/3b-hi-pretrain-research_release
• Finetuned by: parkky21
• Language: Hindi (hi), with Hinglish tolerance
• Voices: अनुष्का (warm, curious), करन (friendly, direct)
• Architecture: LLaMA-family, decoder-only
• Intended use: Multi-turn dialogue in Hindi with lightweight “voice” control via speaker prefixes

---

▶️ Try It (Colab) Use the Colab notebook for inference and examples—no local setup needed: Colab: https://colab.research.google.com/drive/1-greyn4D7-0SVUx86fGPzj5rjB2DjGUn?usp=sharing

---

✨ What’s Special

• Two natural voices out of the box—switch tone by prefixing lines with the speaker name.
• Simple prompting (no special chat template required).
• Fast + lightweight—great for laptops and mid-tier GPUs thanks to Unsloth and 3B size.

---

🗣️ Voices & Prompting

Use speaker-name prefixes followed by a colon. Example conversation style:

अनुष्का: हे करन, क्या आज बारिश ज़्यादा नहीं हो रही?
करन: हाँ, बहुत ज़्यादा! सुबह से रुकने का नाम ही नहीं ले रही।

---

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.