pankajpandey-dev/Qwen3-4B-Hindi-Instruct-v2
pankajpandey-dev/Qwen3-4B-Hindi-Instruct-v2 is a 4 billion parameter instruction-tuned Qwen3 model, specifically fine-tuned to follow instructions and respond naturally in Hindi. Developed by pankajpandey-dev, this compact, bilingual model excels at Hindi instruction-following and is designed for local and edge deployment. It is openly licensed under Apache 2.0, making it suitable for commercial use in Hindi chat and assistant applications.
Loading preview...
Overview
pankajpandey-dev/Qwen3-4B-Hindi-Instruct-v2 is a 4 billion parameter instruction-tuned model based on Qwen3-4B, specifically optimized for the Hindi language. It is part of the Hindi LLM Series, aiming to provide robust Indic-language models for local and edge deployments. The model is fine-tuned on 10,000 curated Hindi instruction-response pairs, enabling strong instruction-following capabilities in Hindi.
Key Capabilities
- Strong Hindi Instruction-Following: Trained on a dedicated dataset to understand and respond to instructions in Hindi.
- Bilingual Support: Handles both Hindi (Devanagari script) and English.
- Compact Size: With approximately 4 billion parameters, it can run efficiently on consumer GPUs and quantizes well for CPU deployment.
- Open License: Released under Apache 2.0, allowing for commercial use.
Training Details
The model was fine-tuned using LoRA (r=32, α=32) via Unsloth, with the training data filtered to ensure genuinely Hindi responses. The resulting LoRA was merged into 16-bit weights.
Intended Use & Limitations
This model is intended for Hindi chat and assistant applications, instruction-following tasks, and general Indic-language experimentation, particularly for local and edge deployments. As a 4B model, it may exhibit factual errors or inconsistencies on complex reasoning tasks and inherits biases from its base model and training data. Validation of outputs is recommended for production use.