Kanha-AI/kanha-kanha.ai-1.7b-grounded-qlora-2ep-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

The Kanha-AI/kanha-kanha.ai-1.7b-grounded-qlora-2ep-v1 is a 1.7 billion parameter language model developed by Kanha-AI, fine-tuned using QLoRA on a Qwen3-1.7B base. This model is specifically designed for grounded question answering, requiring retrieved source context for inference. It excels at website question answering and is intended for research comparing training methods on Kanha website-derived datasets.

Loading preview...

Kanha-AI/kanha-kanha.ai-1.7b-grounded-qlora-2ep-v1 Overview

This model is a 1.7 billion parameter language model developed by Kanha-AI, built upon the Qwen/Qwen3-1.7B base model. It was fine-tuned using the QLoRA method over 2 epochs, specifically for grounded inference tasks. The model's core design mandates the provision of retrieved source context for any query, making it specialized for scenarios where answers must be derived strictly from provided information.

Key Capabilities & Features

  • Grounded Question Answering: Designed to answer questions exclusively from supplied context, refusing to answer if information is absent.
  • Contextual Integrity: Utilizes a strict prompt contract, including a system prompt that enforces answering "only from the supplied context" and a specific refusal string: "I can't answer that from the provided context."
  • Efficient Fine-tuning: Trained with QLoRA, employing hyperparameters such as a maximum sequence length of 2048 and a learning rate of 0.0001.
  • Evaluation Metrics: Achieved high recall rates on dates_recall (1.0), numbers_recall (0.97), and urls_recall (1.0), with a refusal_rate of 0.115.
  • MLC Availability: Provides MLC artifacts with q4f16_1 quantization for optimized deployment.

Intended Use Cases

  • Research: Primarily intended for research purposes, particularly for comparing training methods on Kanha website-derived datasets.
  • Controlled Evaluation: Suitable for controlled evaluation of website question answering systems where answers must be strictly grounded in provided text.

Limitations

Users should be aware that the model may produce incorrect, incomplete, or stale answers and might memorize training content. It is crucial to review outputs and test representative failure cases before any user-facing deployment.

Finetunes
1 models