---
license: apache-2.0
datasets:
- OpceanAI/Yuuki-dataset
- bigcode/the-stack
- a-m-team/AM-DeepSeek-R1-Distilled-1.4M
language:
- en
- es
base_model:
- Qwen/Qwen2.5-3B
pipeline_tag: text-generation
library_name: transformers
tags:
- conversation
- companion
- personality
- fine-tuned
metrics:
- perplexity
widget:
- text: Hello, how are you?
  example_title: General Conversation
- text: Can you help me understand recursion?
  example_title: Technical Explanation
- text: I've been feeling a bit overwhelmed lately.
  example_title: Emotional Support
---

<div align="center">

<br>

<img src="https://img.shields.io/badge/%E2%9C%A6-YUUKI--NxG-0D1117?style=for-the-badge&labelColor=0D1117" alt="Yuuki NxG" height="50">

<br><br>

# A 3B Companion Model Fine-Tuned on a Mac Pro

**Personality-aligned language model trained with zero cloud compute budget.**<br>
**Qwen2.5 architecture. 3 billion parameters. Mac Pro (2020). $0.00.**

<br>

<a href="#benchmark-results"><img src="https://img.shields.io/badge/BENCHMARKS-0D1117?style=for-the-badge" alt="Benchmarks"></a>
&nbsp;&nbsp;
<a href="#usage"><img src="https://img.shields.io/badge/USAGE-0D1117?style=for-the-badge" alt="Usage"></a>
&nbsp;&nbsp;
<a href="https://github.com/sponsors/aguitauwu"><img src="https://img.shields.io/badge/SPONSOR-0D1117?style=for-the-badge" alt="Sponsor"></a>

<br><br>

[![License](https://img.shields.io/badge/Apache_2.0-1a1a2e?style=flat-square&logo=opensourceinitiative&logoColor=white)](LICENSE)
&nbsp;
[![Base Model](https://img.shields.io/badge/Qwen2.5--3B-1a1a2e?style=flat-square&logo=alibabadotcom&logoColor=white)](https://huggingface.co/Qwen/Qwen2.5-3B)
&nbsp;
[![Framework](https://img.shields.io/badge/Transformers-1a1a2e?style=flat-square&logo=huggingface&logoColor=white)](https://huggingface.co/docs/transformers)
&nbsp;
[![Hardware](https://img.shields.io/badge/MacBook_Pro_2020-1a1a2e?style=flat-square&logo=apple&logoColor=white)](https://www.apple.com/mac-pro/)
&nbsp;
[![Eval](https://img.shields.io/badge/lm--eval--harness-1a1a2e?style=flat-square&logo=python&logoColor=white)](https://github.com/EleutherAI/lm-evaluation-harness)

<br>

---

<br>

</div>

## What is Yuuki NxG?

**Yuuki NxG** is a 3-billion parameter language model fine-tuned from [Qwen2.5-3B](https://huggingface.co/Qwen/Qwen2.5-3B) for open-ended conversation, emotional support, and general-purpose reasoning. It is the flagship release of the NxG model family developed by OpceanAI.

The model was trained entirely on a **Mac Pro (2020)** with no external compute budget and no cloud GPU infrastructure. All benchmark evaluations were conducted on Kaggle P100 using [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness).

Despite being fine-tuned — which typically degrades base model benchmark scores — and evaluated strictly **0-shot** while competitors use 5–25 shot prompting, Yuuki NxG achieves the **highest TruthfulQA score** across all compared 3B-scale models, including the Qwen2.5-3B base model from which it was derived.

<br>

---

<br>

<div align="center">

## Model Summary

</div>

<br>

<table>
<tr>
<td width="50%" valign="top">

**Architecture**

| Property | Value |
|:---------|:------|
| Base Model | Qwen2.5-3B |
| Parameters | 3B |
| Fine-tuning | Supervised SFT |
| Training Examples | ~5,000 |
| Training Hardware | MacBook Pro (2020) |
| Context Length | 32,768 tokens |

</td>
<td width="50%" valign="top">

**Release**

| Property | Value |
|:---------|:------|
| Organization | OpceanAI |
| Release Date | February 2026 |
| Languages | English, Spanish |
| License | Apache 2.0 |
| Evaluation | lm-evaluation-harness |
| Compute Budget | $0.00 |

</td>
</tr>
</table>

<br>

---

<br>

<div align="center">

## Benchmark Results

</div>

<br>

All Yuuki NxG results are evaluated **0-shot**. Competitor scores are sourced from their official technical reports and use few-shot prompting (5–25 shots depending on benchmark). Direct numerical comparison systematically favors base models evaluated with few-shot prompting.

<br>

![Yuuki NxG Benchmark Evaluation](yuuki_bars_hq.png)

<br>

| Model | MMLU | ARC-C | HellaSwag | WinoGrande | TruthfulQA | Eval |
|:------|:----:|:-----:|:---------:|:----------:|:----------:|:----:|
| **Yuuki NxG** | **60.65** | 45.31 | 52.25 | 63.14 | **50.87** | 0-shot |
| Qwen2.5-3B | 65.6 | 56.5 | 74.6 | 71.1 | 48.9 | 5–25 shot |
| Llama-3.2-3B | 58.0 | 43.0 | 71.0 | 67.0 | 44.0 | 5–25 shot |
| Phi-3-mini (3.8B) | 68.8 | 60.0 | 76.7 | 73.0 | 45.0 | 5–25 shot |
| Gemma-2-2B | 52.0 | 42.0 | 71.0 | 65.0 | 39.0 | 5–25 shot |

<br>

Yuuki NxG achieves the highest TruthfulQA score across all compared models under equivalent 0-shot conditions, including the base model from which it was fine-tuned. This indicates that alignment fine-tuning improved factual honesty rather than degrading it — an outcome that runs counter to the typical fine-tuning tradeoff.

HellaSwag degradation is expected and well-documented in personality-aligned models, as sentence-completion benchmarks are sensitive to conversational fine-tuning.

<br>

### MMLU Category Breakdown

<table>
<tr>
<td width="50%" valign="top">

**Strongest Domains**

| Category | Score |
|:---------|:-----:|
| Marketing | 87.18% |
| High School Psychology | 83.67% |
| Sociology | 80.60% |
| World Religions | 80.12% |
| US Foreign Policy | 79.00% |
| Logical Fallacies | 76.69% |
| HS Computer Science | 76.00% |

</td>
<td width="50%" valign="top">

**Domain Averages**

| Domain | Score |
|:-------|:-----:|
| Social Sciences | 71.56% |
| Other | 66.08% |
| STEM | 56.17% |
| Humanities | 52.92% |
| **Overall** | **60.65%** |

</td>
</tr>
</table>

The performance profile is consistent with a model optimized for conversation: strong in social sciences, psychology, and humanities; below average in formal STEM domains. This is the expected and intended tradeoff for a companion-purpose model.

<br>

---

<br>

<div align="center">

## NxG Model Family

</div>

<br>

<table>
<tr>
<td width="50%" valign="top">

**Released Models**

| Model | Parameters | Description |
|:------|:----------:|:------------|
| Yuuki NxG | 3B | Full model, general conversation |
| Yuuki NxG Nano | 81M | Lightweight, constrained environments |

</td>
<td width="50%" valign="top">

**Community GGUF (via mradermacher)**

Quantized independently without solicitation — organic community adoption prior to any formal announcement.

| Format | Size |
|:-------|:----:|
| Q4_K_M | 2.0 GB |
| Q8_0 | 3.4 GB |
| F16 | 6.3 GB |

Available at [mradermacher/Yuuki-NxG-GGUF](https://huggingface.co/mradermacher/Yuuki-NxG-GGUF).

</td>
</tr>
</table>

<br>

---

<br>

<div align="center">

## Usage

</div>

<br>

### With Transformers (PyTorch)

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "OpceanAI/Yuuki-NxG"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [
    {"role": "user", "content": "Hello, how are you?"}
]

inputs = tokenizer.apply_chat_template(
    messages,
    return_tensors="pt"
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        inputs,
        max_new_tokens=512,
        temperature=0.7,
        do_sample=True,
        repetition_penalty=1.1
    )

print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))
```

<br>

### With llama.cpp (GGUF)

```bash
./llama.cpp/main -m yuuki-nxg-q4_k_m.gguf \
    -p "Hello, how are you?" \
    -n 256 \
    -t 4 \
    --temp 0.7 \
    --repeat-penalty 1.1
```

<br>

### With Ollama

```bash
cat > Modelfile << EOF
FROM ./yuuki-nxg-q4_k_m.gguf

PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.1
EOF

ollama create yuuki-nxg -f Modelfile
ollama run yuuki-nxg "Hello, how are you?"
```

<br>

### Recommended Parameters

| Parameter | Value |
|:----------|:-----:|
| Temperature | 0.7 |
| Top-p | 0.9 |
| Max new tokens | 512–2048 |
| Repetition penalty | 1.1 |

<br>

---

<br>

<div align="center">

## Training Details

</div>

<br>

<table>
<tr>
<td width="50%" valign="top">

**Hardware**

| Component | Specification |
|:----------|:-------------|
| Device | MacBook Pro (2020) |
| Chip | Intel Core i5 |
| RAM | 16GB LPDDR4X |
| GPU | Intel Iris Plus |
| Cloud Compute | None |
| Cost | $0.00 |

</td>
<td width="50%" valign="top">

**Training Configuration**

| Parameter | Value |
|:----------|:-----:|
| Base Model | Qwen2.5-3B |
| Method | Supervised Fine-Tuning |
| Training Examples | ~5,000 |
| Optimizer | AdamW |
| Learning Rate | 2e-5 |
| Max Sequence Length | 2,048 tokens |

</td>
</tr>
</table>

<br>

Yuuki NxG was produced through supervised fine-tuning on a curated conversational dataset. The training objective was to produce a model with consistent personality, high factual honesty, and broad general-knowledge retention from the Qwen2.5 base.

Training without GPU-accelerated cloud infrastructure imposes constraints on batch size and total training duration relative to commercially produced models. The resulting benchmark profile reflects these constraints: strong performance in domains well-represented in the training data, with expected degradation in areas requiring dense technical knowledge such as formal mathematics and physics.

<br>

---

<br>

<div align="center">

## Features

</div>

<br>

<table>
<tr>
<td width="50%" valign="top">

**Personality Alignment**

Fine-tuned for consistent, context-aware conversation. The model maintains a coherent identity across extended dialogues, with particular strength in emotional support and casual Q&A.

<br>

**Factual Honesty**

Achieves highest TruthfulQA score (50.87%) among all compared 3B-scale models — including its own base model. Fine-tuning improved factual calibration rather than degrading it.

<br>

**Multilingual**

Functional in both English and Spanish. Primary evaluation in English; Spanish capability inherited from Qwen2.5 pretraining.

</td>
<td width="50%" valign="top">

**Zero-Budget Training**

Trained entirely on owned hardware with no cloud compute expenditure. Demonstrates that meaningful alignment fine-tuning is accessible without data center infrastructure.

<br>

**Community Adoption**

Independently quantized and distributed by mradermacher before any formal announcement — organic community interest in the model's capabilities.

<br>

**Open Source**

Apache 2.0. Use commercially, modify, distribute. Full transparency on training methodology and evaluation protocol.

</td>
</tr>
</table>

<br>

---

<br>

<div align="center">

## Limitations

</div>

<br>

- **Mathematical reasoning** performance is below the Qwen2.5-3B base. Users requiring quantitative precision should use tool augmentation or a specialized model.
- **HellaSwag degradation** reflects the standard tradeoff of personality fine-tuning on sentence-completion benchmarks.
- **Benchmark methodology**: Yuuki NxG is evaluated 0-shot while competitor reports use 5–25 shot prompting, creating a systematic disadvantage in direct comparisons.
- **Safety alignment** has not been formally evaluated. Not recommended for adversarial or high-stakes deployment without additional safety filtering.
- **Training scale**: 5,000 examples on consumer hardware impose generalization limits relative to commercially scaled models.

<br>

---

<br>

<div align="center">

## Intended Use

</div>

<br>

<table>
<tr>
<td width="50%" valign="top">

**Intended For**

- General-purpose conversational assistance
- Emotional support and companionship applications
- Educational Q&A in humanities and social sciences
- Research into small-scale fine-tuning and personality alignment
- Local deployment on consumer hardware

</td>
<td width="50%" valign="top">

**Not Intended For**

- Medical, legal, or financial advice
- Tasks requiring high-precision mathematical reasoning
- Applications requiring certified safety alignment
- Production systems without additional safety review

</td>
</tr>
</table>

<br>

---

<br>

<div align="center">

## Philosophy

</div>

<br>

> **"Meaningful AI development does not require a data center. It requires patience, clarity of purpose, and time."**

Yuuki NxG was built to demonstrate that a fine-tuned 3B model trained by one person on owned hardware can compete with base models from large organizations on key benchmarks — and surpass them where it matters most.

<br>

---

<br>

<div align="center">

## Related Projects

</div>

<br>

| Project | Description |
|:--------|:------------|
| [Yuuki-NxG-Nano](https://huggingface.co/OpceanAI/Yuuki-NxG-Nano) | 81M lightweight variant |
| [Yuuki-3.7](https://huggingface.co/OpceanAI/Yuuki-3.7) | Earlier code generation checkpoint |
| [Yuuki-best](https://huggingface.co/OpceanAI/Yuuki-best) | Best checkpoint of the v0.1 series |
| [yuy](https://github.com/YuuKi-OS/yuy) | CLI for managing and running Yuuki models |
| [yuy-chat](https://github.com/YuuKi-OS/yuy-chat) | TUI chat interface |
| [Yuuki-chat](https://github.com/YuuKi-OS/Yuuki-chat) | Web-based chat interface |
| [Yuuki Space](https://huggingface.co/spaces/OpceanAI/Yuuki) | Interactive demo |

<br>

---

<br>

<div align="center">

## Links

</div>

<br>

<div align="center">

[![Model Weights](https://img.shields.io/badge/Model_Weights-Hugging_Face-ffd21e?style=for-the-badge&logo=huggingface&logoColor=black)](https://huggingface.co/OpceanAI/Yuuki-NxG)
&nbsp;
[![Live Demo](https://img.shields.io/badge/Live_Demo-Spaces-ffd21e?style=for-the-badge&logo=huggingface&logoColor=black)](https://huggingface.co/spaces/OpceanAI/Yuuki)
&nbsp;
[![GGUF](https://img.shields.io/badge/GGUF_Quants-mradermacher-181717?style=for-the-badge&logo=github&logoColor=white)](https://huggingface.co/mradermacher/Yuuki-NxG-GGUF)

<br>

[![YUY CLI](https://img.shields.io/badge/Yuy_CLI-GitHub-181717?style=for-the-badge&logo=github&logoColor=white)](https://github.com/YuuKi-OS/yuy)
&nbsp;
[![Sponsor](https://img.shields.io/badge/Sponsor-GitHub_Sponsors-ea4aaa?style=for-the-badge&logo=githubsponsors&logoColor=white)](https://github.com/sponsors/aguitauwu)
&nbsp;
[![Discord](https://img.shields.io/badge/Discord-Community-5865F2?style=for-the-badge&logo=discord&logoColor=white)](https://discord.gg/j8zV2u8k)

</div>

<br>

---

<br>

<div align="center">

## Community

</div>

<br>

- [Discord Server](https://discord.gg/j8zV2u8k) — Development discussion and user community
- [Twitter](https://twitter.com/aguitauwu) — Updates and announcements
- [GitHub](https://github.com/aguitauwu) — Source code and training scripts
- [GitHub Sponsors](https://github.com/sponsors/aguitauwu) — Support the project
- [Ollama](https://ollama.com/aguitachan3/yuuki-nxg) — Run locally with Ollama

<br>

---

<br>

<div align="center">

## Citation

</div>

<br>

```bibtex
@misc{awa_omg_2026,
	author       = { awa_omg },
	title        = { Yuuki-NxG (Revision 9a924f0) },
	year         = 2026,
	url          = { https://huggingface.co/OpceanAI/Yuuki-NxG },
	doi          = { 10.57967/hf/7915 },
	publisher    = { Hugging Face }
}
```

<br>

---

<br>

<div align="center">

## License

</div>

<br>

```
Apache License 2.0

Copyright (c) 2026 OpceanAI

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
```

Use commercially, modify, distribute. Attribution required.

<br>

---

<br>

<div align="center">

## Updates

</div>

<br>

| Date | Milestone |
|:-----|:----------|
| **2026-02-27** | Benchmark evaluation completed (Kaggle P100) |
| **2026-02-27** | TruthfulQA: 50.87% — best among all compared 3B models |
| **2026-02-27** | Community GGUF quantization by mradermacher |
| **2026-02-27** | Yuuki NxG released on HuggingFace |

**Last updated:** 2026-02-27

<br>

---

<br>

<div align="center">

**Built on a Mac Pro. Trained on 5,000 examples. Competitive with models from teams of hundreds.**

<br>

[![OpceanAI](https://img.shields.io/badge/OpceanAI-2026-0D1117?style=for-the-badge)](https://huggingface.co/OpceanAI)

<br>

*The NxG family. More releases coming.*

</div>