sugiv/qwen3-8b-tanglish

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The sugiv/qwen3-8b-tanglish model is an 8 billion parameter Qwen3-8B fine-tune, specifically optimized to reply in casual Tanglish (code-mixed Tamil transliterated into the Latin alphabet). It was trained on 81,261 SFT examples from the sugiv/tanglish-pairs-v1 dataset using LoRA, achieving superior authenticity and helpfulness in Tanglish conversations compared to the base model. This model excels at generating natural, code-mixed responses for applications targeting South Indian linguistic contexts, with a 32768 token context length.

Loading preview...

Model Overview

sugiv/qwen3-8b-tanglish is a specialized fine-tune of the 8 billion parameter Qwen3-8B model, developed by sugiv. Its primary purpose is to generate responses in casual Tanglish, which is a code-mixed form of Tamil transliterated into the Latin alphabet, commonly spoken in Chennai and across South India. The model was trained using LoRA (r=16, alpha=32) on the sugiv/tanglish-pairs-v1 dataset, comprising 81,261 supervised fine-tuning examples.

Key Capabilities & Performance

  • Authentic Tanglish Generation: Significantly outperforms the stock Qwen3-8B on LLM-judge dimensions for Tanglish authenticity, intelligibility, and naturalness in both single-turn and multi-turn prompts.
  • Faster Inference: Achieves approximately 7.5x faster inference than the base model by suppressing verbose internal reasoning (<think>...</think>) for casual chat scenarios.
  • Zero Tamil-script Leaks: Successfully avoids generating Tamil script, focusing purely on Latin-script Tanglish.
  • Robust Training: Trained in bf16 precision on an L40S 48 GB GPU, ensuring cleaner training compared to QLoRA/4-bit methods.

Ideal Use Cases

  • Conversational AI: Building chatbots or virtual assistants that interact naturally in Tanglish.
  • Content Generation: Creating text content, dialogues, or social media posts in code-mixed Tamil and English.
  • Regional Applications: Developing applications specifically for users in South India who communicate in Tanglish.

This model is particularly suited for scenarios where authentic, casual Tanglish communication is crucial, offering a distinct advantage over general-purpose LLMs.

Finetunes
1 models