Kanha-AI/kanha-kanha.ai-1.7b-qlora-quality
Kanha-AI/kanha-kanha.ai-1.7b-qlora-quality is a 1.7 billion parameter language model developed by Kanha-AI, based on the Qwen3 architecture. This model was fine-tuned using QLoRA on a Kanha website-derived dataset, focusing on website question answering. It is intended for research into training methods and controlled evaluation of its question-answering capabilities. The model has a maximum sequence length of 2048 tokens and is available in bfloat16 dtype.
Loading preview...
Kanha-AI/kanha-kanha.ai-1.7b-qlora-quality Overview
This model is a 1.7 billion parameter language model from Kanha-AI, built upon the Qwen/Qwen3-1.7B base architecture. It was fine-tuned using the QLoRA method with a maximum sequence length of 2048 tokens, specifically on a dataset derived from the Kanha website. The final merged model is in bfloat16 dtype.
Key Characteristics & Evaluation
The model's training involved 20 epochs with a learning rate of 0.0002 and a LoRA rank of 64. Evaluation metrics indicate strong performance in specific recall tasks, achieving 1.0 for dates_recall and urls_recall, and 0.824 for numbers_recall. However, deterministic_pass_rate is 0.038 and list_recall is 0.127.
Intended Use & Limitations
This checkpoint is primarily intended for research purposes, specifically for comparing different training methods on the same Kanha website-derived dataset and for controlled evaluation of website question answering. It is important to note that the model may produce incorrect, incomplete, or stale answers and can memorize training content. Users should validate outputs and test failure cases before any user-facing deployment. MLC artifacts with q4f16_1 quantization are available for deployment.