laabamone/laabam-ai-3b-v1
laabamone/laabam-ai-3b-v1 is a multilingual AI assistant fine-tuned from Qwen2.5-3B-Instruct using QLoRA. This 3 billion parameter model is optimized for general instruction following, coding, reasoning, and safety alignment across English and several Indic languages including Hindi, Telugu, Kannada, and Tamil. It is designed for developers seeking a compact, instruction-tuned model with strong multilingual capabilities, particularly for South Asian languages.
Loading preview...
Laabam AI 3B v1: Multilingual Instruction Follower
Laabam AI 3B v1 is a compact, multilingual AI assistant developed by laabamone, fine-tuned from the Qwen2.5-3B-Instruct base model. It leverages QLoRA (r=16, alpha=32) for efficient training, making it suitable for a range of instruction-following tasks.
Key Capabilities
- Multilingual Support: Proficient in English, Hindi, Telugu, Kannada, and Tamil, making it valuable for applications targeting South Asian linguistic contexts.
- Instruction Following: Trained for general instruction adherence, covering diverse prompts.
- Reasoning & Coding: Demonstrates capabilities in logical reasoning and code generation.
- Safety Alignment: Incorporates safety alignment during its training process to ensure responsible outputs.
Training Details
The model underwent 4 epochs of training on approximately 98K samples, achieving a final training loss of 0.465. The training regimen progressively expanded its focus from core instruction following to include safety alignment and specific Indic languages, utilizing a carefully managed learning rate schedule to refine performance and prevent catastrophic forgetting.
Good For
- Applications requiring a small, efficient instruction-tuned model.
- Use cases involving multilingual interactions, especially with English and the specified Indic languages.
- Tasks that benefit from general instruction following, basic coding, and reasoning capabilities.