pankajpandey-dev/qwen3.5-9b-hindi-instruct
pankajpandey-dev/qwen3.5-9b-hindi-instruct is a 9 billion parameter instruction-tuned causal language model, fine-tuned from unsloth/Qwen3.5-9B. Developed by pankajpandey-dev, this model is specifically optimized to generate fluent, native Hindi responses without English code-switching or internal English 'thinking' detours. It excels at direct Hindi text generation and is suitable for applications requiring pure Devanagari output, even running on laptop CPUs via its GGUF version.
Loading preview...
Overview
pankajpandey-dev/qwen3.5-9b-hindi-instruct is a 9 billion parameter language model, fine-tuned from the Qwen3.5-9B base model. Its primary distinction is its ability to generate fluent, native Hindi responses without the common issue of English code-switching or internal English 'thinking' processes often observed in general-purpose multilingual models. This fine-tune ensures direct and pure Devanagari output, making it highly effective for Hindi-specific applications.
Key Capabilities
- Native Hindi Generation: Produces responses exclusively in Hindi, avoiding English interjections or internal processing.
- Instruction Following: Tightly adheres to instructions provided in Hindi.
- Efficiency: Designed to run efficiently, with a GGUF version available for laptop CPU deployment (~5.7 GB).
- Multimodal Base (Text-only use): While based on a multimodal Qwen3.5-9B, this fine-tune is optimized for text generation.
Training Details
The model was fine-tuned using LoRA (r=16, alpha=16) with response-only loss on a dataset of 12,912 Hindi pairs. This dataset comprised diverse sources including anudesh (5,000 pairs), dolly-hi (4,000 pairs), wikiHow-hi (3,000 pairs), and Aya-hi (912 pairs). Training involved 2 epochs with a learning rate of 1e-4 cosine schedule, an effective batch size of 16, and a sequence length of 2048, completed on a single NVIDIA L40S GPU in approximately 135 minutes.
Limitations
- Some training data is machine-translated or model-generated, which may lead to occasional unnatural phrasing or factual inaccuracies.
- Longer outputs, such as formal letters, may exhibit repetition without the application of
repetition_penalty=1.1. - The model has not undergone additional safety tuning.
- Knowledge cutoff aligns with the base Qwen3.5 model.
Good for
- Applications requiring direct and unadulterated Hindi text generation.
- Developing chatbots or content generation tools specifically for Hindi speakers.
- Use cases where avoiding English code-switching is critical for user experience.
- Deployment on resource-constrained devices like laptops via its GGUF variant.