tufanakbas23/Pusula-danisman-ai

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 19, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Pusula-danisman-ai by tufanakbas23 is a 12 billion parameter language model, fine-tuned from Mistral-Nemo-Instruct-2407 with a 32768 token context length. It is specifically optimized for Turkish-language guidance on TEKNOFEST and TÜBİTAK project processes, including idea clarification, technical report writing, and jury preparation. This model excels at providing context-aware assistance for project development within these specific Turkish competition frameworks.

Loading preview...

Pusula-danisman-ai: Turkish AI Consultant for TEKNOFEST and TÜBİTAK

Pusula-danisman-ai is a specialized 12 billion parameter language model developed by tufanakbas23, fine-tuned from mistralai/Mistral-Nemo-Instruct-2407. Unlike general-purpose LLMs, this model is specifically configured to assist participants in TEKNOFEST and TÜBİTAK competitions, addressing the complexities of project application processes in Turkey.

Key Capabilities

  • Project Idea Refinement: Helps clarify vague project ideas and narrow their scope.
  • Technical Report Writing: Supports drafting problem definitions, objectives, methodologies, and unique value propositions for KTR/PTR reports.
  • Application Guidance: Provides direction aligned with TÜBİTAK application logic.
  • Analysis Support: Assists with applicability and risk analysis.
  • Content Generation: Generates titles, summaries, and abstracts.
  • Jury Preparation: Aids in preparing for presentations, pitches, and jury evaluations.

Training and Specialization

The model was fine-tuned using QLoRA (4-bit) with unsloth on a unique dataset of 3043 Turkish instruction-output pairs. This dataset is highly specialized, covering categories like report writing, project development from scratch, general summaries, strategy, and error correction. A notable aspect of its training includes a 'red' category, teaching the model to explicitly reject out-of-domain questions (e.g., about food, stock market, health).

System Architecture

Pusula-danisman-ai operates within a three-layered architecture, including data/training, behavior, and application layers. The behavior layer uses system prompts, routing guides, and defined 'skills' (e.g., report_section_writer, project_idea_refinement) to ensure the model provides contextually appropriate and task-oriented responses, rather than just generic text.

Limitations

  • Responds only in Turkish.
  • Explicitly rejects out-of-domain questions.
  • Performance may degrade for engineering topics outside the TEKNOFEST/TÜBİTAK context.

Finetunes
1 models