Kanha-AI/kanha-kanha.ai-1.7b-grounded-qlora-v1

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 22, 2026Architecture:Transformer Featherless Exclusive Cold

Kanha-AI/kanha-kanha.ai-1.7b-grounded-qlora-v1 is a 1.7 billion parameter language model developed by Kanha-AI, fine-tuned using QLoRA on the Qwen3-1.7B base model. It is specifically designed for grounded inference, requiring a supplied context to answer questions, and excels at website question answering. The model has a context length of 2048 tokens and is intended for research into training methods and controlled evaluation of website Q&A.

Loading preview...

Kanha-AI/kanha-kanha.ai-1.7b-grounded-qlora-v1 Overview

This model is an experimental 1.7 billion parameter language model developed by Kanha-AI, based on the Qwen/Qwen3-1.7B architecture. It was fine-tuned using the QLoRA method with a maximum sequence length of 2048 tokens. A key characteristic is its grounded inference contract, meaning it is designed to answer questions only from provided source context. Queries without context are outside its trained scope.

Key Capabilities and Training

  • Grounded Question Answering: Optimized to provide concise answers strictly from supplied context, refusing to answer if information is absent.
  • Website Q&A: Specifically trained and evaluated for question answering based on website-derived datasets.
  • QLoRA Fine-tuning: Utilizes QLoRA with specific hyperparameters (rank 16, alpha 16, dropout 0.05) targeting various projection layers.
  • Evaluation Metrics: Achieved high recall rates on dates_recall (1.0) and urls_recall (1.0), with a refusal_rate of 0.15.

Intended Use and Limitations

This model is primarily intended for research purposes, specifically for comparing training methods on Kanha's website-derived datasets and for controlled evaluation of website question answering systems. Users should be aware that the model may produce incorrect, incomplete, or stale answers, and can memorize training content. It is crucial to review outputs and qualify the model in target environments before any user-facing deployment.

Finetunes
1 models