Kanha-AI/kanha-kanha.ai-1.7b-full
Kanha-AI/kanha-kanha.ai-1.7b-full is a 1.7 billion parameter language model developed by Kanha-AI, based on the Qwen3 architecture. This model is an experimental full fine-tune, specifically trained on a Kanha website-derived dataset. It is intended for research into training methods and controlled evaluation of website question answering, particularly for recalling dates, URLs, and numbers.
Loading preview...
Kanha-AI/kanha-kanha.ai-1.7b-full: Qwen3 Experiment
This model, kanha.ai-1.7b-full, is an experimental full fine-tune of the Qwen/Qwen3-1.7B base model, developed by Kanha-AI. It was trained using a specific Kanha website-derived dataset, comprising 128 training records and 24 validation records, with a maximum sequence length of 2048 tokens.
Key Characteristics & Evaluation
The model demonstrates strong performance in specific recall tasks:
- Dates Recall: Achieved a perfect score of
1.0. - URLs Recall: Also achieved a perfect score of
1.0. - Numbers Recall: Scored
0.75.
However, it showed a low list_recall at 0.02 and an unsupported_value_rate of 0.38. The training utilized a bfloat16 dtype and an assistant-only loss function.
Intended Use Cases
This checkpoint is primarily intended for:
- Research: Comparing different training methods on the same Kanha website-derived dataset.
- Controlled Evaluation: Assessing its performance in website question answering scenarios.
Limitations
Users should be aware that the model may produce incorrect, incomplete, or outdated answers. It also has the potential to memorize training content. It is crucial to review outputs, test for failure cases, and qualify the exact runtime environment before any user-facing deployment.