Model Overview: DesiFamily-2B
The suneeldk/DesiFamily-2B model is a 2.6 billion parameter language model, developed by suneeldk, specifically fine-tuned to generate humorous, character-driven text in the style of various Indian family members. Built upon the Gemma-2-2B-it base, it leverages Unsloth and LoRA for efficient fine-tuning.
Key Capabilities
- Character-Specific Roasting: The model can adopt the persona of several distinct Indian family archetypes, including:
- Indian Mom: Guilt trips and emotional blackmail.
- Indian Dad: Silent disappointment and salary comparisons.
- Jealous Aunty: Delivers backhanded compliments and shade.
- Random Uncle: References "my friend's son" for comparison.
- Pure English Output: All generated responses are in pure English, avoiding Hinglish or other mixed languages, making it accessible for a broad audience.
- Humorous Text Generation: Optimized for creating comedic and relatable "roasts" based on common family interactions.
Training and Specifications
- Base Model: Gemma-2-2B-it.
- Fine-tuning: Utilized Unsloth and LoRA.
- Dataset: Trained on over 1200 hand-crafted, pure-English roasting examples.
- Context Length: Supports an 8192-token context window.
- License: Apache 2.0, allowing for broad usage.
Ideal Use Cases
This model is best suited for:
- Meme Generation: Creating humorous content and social media posts.
- Creative Writing: Developing character dialogue for comedic sketches or stories.
- Entertainment: Generating amusing and relatable family-style interactions for digital platforms.