large-traversaal/Qwen-2.5-14B-Hindi

TEXT GENERATIONConcurrency Cost:1Model Size:14.8BQuant:FP8Ctx Length:32kPublished:Feb 27, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Qwen-2.5-14B-Hindi is a 14.8 billion parameter instruction-tuned bilingual large language model developed by Traversaal.ai and 1-800-LLMs. Optimized for both Hindi and English, it demonstrates improved performance on Hindi tasks (approx. 3.5% better) and tougher English benchmarks (approx. 7.6% better) compared to the original Qwen-2.5-14B-Instruct. This model is designed for bilingual, multilingual, and non-English applications, serving as a foundational model for fine-tuning in areas like AI chat assistants and customer support.

Loading preview...

Model Overview

Qwen-2.5-14B-Hindi is a 14.8 billion parameter instruction-tuned large language model developed by Traversaal.ai and 1-800-LLMs. It is specifically optimized for bilingual performance in Hindi and English, showing notable improvements over the original Qwen-2.5-14B-Instruct model.

Key Capabilities & Performance

  • Enhanced Hindi Performance: Achieves approximately 3.5% better average performance on Hindi tasks compared to the original model.
  • Improved English Benchmarks: Demonstrates around 7.6% better performance on challenging English benchmarks like MMLU-Pro, MATH-Hard, and GPQA.
  • Reduced Bias: Features less bias towards the ordering of choices in multiple-choice questions, a result of modified training techniques.
  • Flexible Prompt Formats: Supports various prompt formats for tasks such as NLI, MCQ, summarization, coding, and translation.

Intended Use Cases

  • Bilingual Applications: Ideal for developing AI-powered chat assistants, customer support services, and educational tools catering to Hindi and English speakers.
  • Research & Development: Serves as a valuable tool for researchers and developers in NLP, particularly for multilingual advancements.
  • Commercial Fine-tuning: Can be utilized as a foundational model for fine-tuning to meet specific industry needs in bilingual contexts.

Limitations

Users should be aware that the model is limited to Hindi and English, may generate incorrect or offensive information, and should not be used for high-risk decision-making without human oversight.

Finetunes
1 models