DoctorEdoP369/Atena_Ai

Atena Ai

Unique. Conversational. Deep Reasoning.

A 3B reasoning model that runs entirely on your machine. Small, fast yet powerful model, trained on filtered, high-quality data.

---

Why this model exists

A model trained on high-quality public and custom datasets, designed to help you. Designed for privacy and efficiency, this lightweight model brings high-level reasoning and analytical power to low-spec mobile devices. By running fully offline through compatible local software, it delivers the performance of a much larger model right on your phone—without your data ever leaving your device. Light Uncensored

What it does well

• Deep reasoning. Built on a thinking-mode base, the model produces an explicit <think> chain before answering. You can read how it got there, not just where it landed. Naturally, it is best to use programs that display the answer directly, to avoid wasting time watching how the model "reasons."
• Short & Punchy (Great for a Quick Summary or Card Description) A compact, highly capable AI designed to run offline on low-power mobile devices. It offers heavy-duty reasoning and analysis while keeping your data 100% private with the right local runner.
• Smart. "Smart" Reasoning: Punches above its weight class, delivering deep analysis and reasoning comparable to much larger models.

---

Model details

• Developed by: DoctorEdoP369
• Base model: HuggingFaceTB/SmolLM3-3B
• Architecture: SmolLM3-3B, thinking mode
• Method: LoRA fine-tuning with Unsloth, trained on responses only
• Quantization: Q6_K and Q8, from a bf16 merge
• License: Apache 2.0

Training data

Fine-tuned on a large mixed corpus.

nvidia/Nemotron-Post-Training-Dataset-v2

Used to strengthen multi-turn reasoning and general answer quality.

License/Terms of Use: the dataset carries per-sample license information. It is predominantly CC-BY-4.0, with a small subset of prompts from WildChat under ODC-BY and a small subset from StackOverflow under CC-BY-SA.

This dataset contains synthetic data created using DeepSeek-R1-0528, Qwen2.5-14B-Instruct, Qwen2.5-32B-Instruct-AWQ, Qwen3-30B-A3B and Qwen3-235B-A22B. If this dataset is used to create, train, fine-tune, or otherwise improve an AI model, which is distributed or made available, such AI model may be subject to redistribution and use requirements in the Qwen License Agreement and the DeepSeek License Agreement.

Data Developer: NVIDIA

facebook/natural_reasoning

Used to broaden open-ended reasoning across domains.

Released by Meta under CC-BY-NC-4.0. This is a non-commercial license. Users who intend to deploy this model commercially should review the terms on the dataset page and assess their own position.

The remaining training material consists of custom high-quality datasets built for this project, including original Italian multi-turn reasoning data.

HelioAI/Fable-5-Distill-Reasoning-462x

I took thousands of examples from this dataset, and filtered and improved everything with Opus 4.8

High-quality custom datasets.

example: DoctorEdoP369/bonsai_8b_distilled_edited_106byDoctorEdoP369

---

Limitations and honest notes

• This AI model is powerful, but features reduced security filtering thanks to custom datasets. The impact is very low; it's not an uncensored model.
• Please note that the use of this model including the associated questions and answers is entirely your responsibility.
• Neither the author of this model nor the authors of the models used as a basis (also datasets) for it bear any liability for how you choose to use it.
• The responsibility for what you ask, and what you do with the answer, is entirely yours.

---

AI can make mistakes; it does not replace professional advice.

Created by DoctorEdoP369

Thanks to the HuggingFaceTB community and the creators of the datasets used.