Kanha-AI/kanha-kanha.ai-1.7b-grounded-qlora-3ep-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-3ep-v1 is a 1.7 billion parameter Qwen3-based language model developed by Kanha-AI, fine-tuned using QLoRA. This model is specifically designed for grounded question answering, requiring a supplied context for inference. It excels at information recall from provided text, achieving high scores in dates, numbers, and URLs recall, making it suitable for controlled website question answering and research into training methods.

Loading preview...

Model Overview

Kanha-AI/kanha-kanha.ai-1.7b-grounded-qlora-3ep-v1 is a 1.7 billion parameter model built upon the Qwen/Qwen3-1.7B base model. It has been fine-tuned using the QLoRA method over 3 epochs, with a maximum sequence length of 2048 tokens. The model's primary characteristic is its grounded inference contract, meaning it is designed to answer questions only from a supplied context. A bare question without retrieved source context is considered outside its trained and evaluated contract.

Key Capabilities & Evaluation

This model demonstrates strong performance in context-dependent information retrieval, as indicated by its evaluation metrics:

  • Dates Recall: 1.0
  • Numbers Recall: 0.97179
  • URLs Recall: 1.0
  • List Recall: 0.60512
  • Refusal Rate: 0.0769 (for questions outside context)

The model uses a specific system prompt that enforces answering only from the supplied context and provides a defined refusal string: "I can't answer that from the provided context."

Intended Use Cases

This checkpoint is specifically intended for:

  • Research: Comparing different training methods on Kanha's website-derived dataset.
  • Controlled Evaluation: Assessing website question answering performance in scenarios where answers must be strictly derived from provided text.

Limitations

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, test failure cases, and qualify the exact runtime before any user-facing deployment.

Finetunes
1 models