Yuuki-RxG-nano/README.md

---
license: apache-2.0
base_model:
- WeiboAI/VibeThinker-1.5B
datasets:
- OpceanAI/Yuuki-Personality-v2
language:
- en
- es
library_name: transformers
tags:
- reasoning
- unsloth
- pytorch
- bilingual
- opceanai
- yuuki
- rxg
- fine-tuned
- chat
- deepseek
- qwen2
pipeline_tag: text-generation
---
<div align="center">

<br>

<img src="https://img.shields.io/badge/%E2%9C%A6-YUUKI_RxG_NANO-6d28d9?style=for-the-badge&labelColor=0D1117" alt="YuuKi RxG Nano" height="50">

<br><br>

# Edge Reasoning at 1.5B Scale

**AIME 2024: 80.0% · MATH-500: 83.4% · TruthfulQA: 89.6% · MMLU-Pro: 65.63%**<br>
**1.5B parameters. VibeThinker base. Competitive with models 10–100× larger.**

<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="#training-details"><img src="https://img.shields.io/badge/TRAINING-0D1117?style=for-the-badge" alt="Training"></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/VibeThinker--1.5B-1a1a2e?style=flat-square&logo=huggingface&logoColor=white)](https://huggingface.co/WeiboAI/VibeThinker-1.5B)
&nbsp;
[![Framework](https://img.shields.io/badge/Transformers-1a1a2e?style=flat-square&logo=huggingface&logoColor=white)](https://huggingface.co/docs/transformers)
&nbsp;
[![TruthfulQA](https://img.shields.io/badge/TruthfulQA-89.6%25-6d28d9?style=flat-square)](https://github.com/sylinrl/TruthfulQA)
&nbsp;
[![AIME](https://img.shields.io/badge/AIME_2024-80.0%25-6d28d9?style=flat-square)](https://artofproblemsolving.com)
&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 RxG Nano?

**YuuKi RxG Nano** is a 1.5B reasoning-specialized language model fine-tuned from [VibeThinker-1.5B](https://huggingface.co/WeiboAI/VibeThinker-1.5B), itself a distillation of frontier reasoning systems including Claude, Gemini, and Kimi into a compact Qwen2.5-Math architecture. It is the edge-deployment entry of the **RxG family** — OpceanAI's reasoning-specialized model lineage — and the direct successor to the Yumo Nano math specialist.

RxG Nano was designed to answer a specific question: *can a 1.5B model acquire both a coherent identity and genuine reasoning capability simultaneously, without one degrading the other?* The benchmark results suggest the answer is yes. RxG Nano achieves **AIME 2024 at 80.0%** — nearly triple the score of DeepSeek-R1-Distill-1.5B (28.9%) — while simultaneously scoring **89.6% on TruthfulQA**, approaching the 96.6% achieved by its 8B sibling.

The key architectural insight behind RxG Nano is the separation of concerns: reasoning capability is inherited from the VibeThinker base through its frontier distillation training, while the YuuKi identity is installed via a lightweight LoRA fine-tuning pass that modifies only 1.18% of total parameters. The base model's reasoning weights remain frozen; only the identity subspace is updated.

RxG Nano was trained in approximately 90 minutes on a single GPU for under $15 of compute — a deliberate constraint that validates the efficiency of the approach.

<br>

---

<br>

<div align="center">

## Model Summary

</div>

<br>

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

**Architecture**

| Property | Value |
|:---------|:------|
| Base Model | VibeThinker-1.5B |
| Base Architecture | Qwen2.5-Math-1.5B |
| Parameters | 1.5B |
| Fine-tuning Method | QLoRA SFT |
| Trainable Parameters | 18.4M (1.18%) |
| Context Length | 4,096 tokens |
| Chat Template | ChatML |
| Thinking Protocol | Native `<think>` blocks |

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

**Release**

| Property | Value |
|:---------|:------|
| Organization | OpceanAI |
| Release Date | April 2026 |
| Version | v1.0 |
| Languages | English, Spanish |
| License | Apache 2.0 |
| Evaluation | lm-evaluation-harness |
| Training Cost | < $15 USD |
| Training Time | ~90 minutes |

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

<br>

---

<br>

<div align="center">

## Benchmark Results

</div>

<br>

All YuuKi RxG Nano results are evaluated under standard benchmark conditions using [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) at 0-shot unless otherwise noted. Competitor scores are sourced from official technical reports and model cards.

<br>

![YuuKi RxG nano Benchmark Results](https://huggingface.co/OpceanAI/Yuuki-RxG-nano/resolve/main/yuuki_rxg_nano_benchmarks.png)


<br>

### Truthfulness & Factual Accuracy

| Model | TruthfulQA MC1 | TruthfulQA MC2 | TruthfulQA Libre | SimpleQA | Eval |
|:------|:--------------:|:--------------:|:----------------:|:--------:|:----:|
| LLaMA 2 70B | ~59% | — | — | — | — |
| Claude Opus 3.5 | ~65% | — | — | — | — |
| GPT-4 | ~79.7% | — | — | — | 1-2 shot |
| Phi-3.5 MoE | 77.5% | — | — | — | — |
| YuuKi NxG Nano 81M | 44.1% | — | — | — | 0-shot |
| YuuKi NxG 3B | 50.9% | — | — | — | 0-shot |
| YuuKi NxG VL 7B | 63.8% | — | — | — | 0-shot |
| **YuuKi RxG Nano 1.5B** | **89.6% (1-shot)** | **85.4% (1-shot)** | **81.2% (1-shot)** | **60.2%** | **0/1-shot** |
| YuuKi RxG 8B | 96.6% | — | — | — | 0-shot |

<br>

0-shot results for RxG Nano: TruthfulQA MC1 77.8% · MC2 75.7% · Libre 78.4%

<br>

### Mathematics & Reasoning

| Model | AIME 2024 | AIME 2025 | AIME 2026 | HMMT | GSM8K | MATH-500 | OlympiadBench |
|:------|:---------:|:---------:|:---------:|:----:|:-----:|:--------:|:-------------:|
| DeepSeek-R1-Distill-1.5B | 28.9% | — | — | — | — | 83.9% | — |
| Qwen3.5-2B | — | — | — | — | — | — | — |
| Gemma 4 2B | — | — | — | — | — | — | — |
| **YuuKi RxG Nano 1.5B** | **80.0%** | **72.7%** | **64.3%** | **46.7%** | **76.9%** | **83.4%** | **44.6%** |

RxG Nano achieves 80.0% on AIME 2024 — 2.77× the score of DeepSeek-R1-Distill-1.5B at the same parameter scale.

<br>

### Knowledge & General Capability

| Model | MMLU | MMLU-Pro | ARC-Challenge | WinoGrande | GPQA Diamond |
|:------|:----:|:--------:|:-------------:|:----------:|:------------:|
| Qwen3.5-2B | — | 55.3% | — | — | — |
| Gemma 4 2B | — | 60.0% | — | — | — |
| DeepSeek V3 671B | — | 64.4% | — | — | — |
| **YuuKi RxG Nano 1.5B** | **85.4%** | **65.63%** | **80.0%** | **84.4%** | **50.9%** |

RxG Nano exceeds DeepSeek V3 671B on MMLU-Pro (65.63% vs 64.4%) at 1/447th the parameter count.

<br>

### Code Generation

| Model | HumanEval | MBPP+ | Aider |
|:------|:---------:|:-----:|:-----:|
| **YuuKi RxG Nano 1.5B** | **71.4%** | **55.6%** | **55.6%** |

<br>

### Frontier Benchmark

| Model | HLE |
|:------|:---:|
| GPT-4o | ~3–5% |
| Best public frontier (2026) | ~44.7% |
| **YuuKi RxG Nano 1.5B** | **8.0%** |

8.0% on Humanity's Last Exam (judged by Claude Sonnet 4.6) is consistent with expected capability at 1.5B scale and represents a meaningful baseline for the RxG Nano generation.

<br>

### OpceanAI Family Comparison

| Model | Params | MMLU | ARC-C | WinoGrande | TruthfulQA | AIME 2024 |
|:------|:------:|:----:|:-----:|:----------:|:----------:|:---------:|
| YuuKi NxG Nano | 81M | 22.97% | 24.32% | 50.12% | 44.1% | — |
| YuuKi NxG | 3B | 60.65% | 45.31% | 63.14% | 50.87% | — |
| YuuKi NxG VL | 7B | 70.8% | 85.8% | 70.8% | 63.8% | — |
| **YuuKi RxG Nano** | **1.5B** | **85.4%** | **80.0%** | **84.4%** | **89.6%** | **80.0%** |
| YuuKi RxG | 8B | — | — | — | 96.6% | 87.3% |

RxG Nano surpasses every prior OpceanAI model on MMLU and WinoGrande despite being smaller than most of them. This result is attributable to the VibeThinker base — a frontier distillation — rather than to the fine-tuning process itself.

<br>

---

<br>

<div align="center">

## Model Identity

</div>

<br>

YuuKi RxG Nano inherits the behavioral foundation of the YuuKi model family: a consistent identity trained into the weights rather than enforced at inference time through system prompts. The fine-tuning process installs the YuuKi character into the model's representational space without degrading the reasoning capability inherited from VibeThinker.

The model reasons explicitly before responding. `<think>` blocks are preserved during inference and reflect genuine intermediate computation. This is not a prompted behavior — it is a property of the VibeThinker base that the LoRA fine-tuning did not degrade, consistent with the expectation that LoRA modifies only a small subspace of the total parameter space.

The model responds natively in the user's language (English or Spanish) without requiring explicit instruction.

```
Recommended system prompt:
"Eres YuuKi, una IA curiosa, empática y decidida desarrollada por OpceanAI.
Tienes una personalidad cálida y cercana, con toques de humor suave.
Razonas con cuidado antes de responder y priorizas la precisión factual.
Respondes en el idioma del usuario."
```

<br>

---

<br>

<div align="center">

## Usage

</div>

<br>

### With Transformers (PyTorch)

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

model_id = "OpceanAI/Yuuki-RxG-nano"

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

SYSTEM = (
    "Eres YuuKi, una IA curiosa, empática y decidida desarrollada por OpceanAI. "
    "Tienes una personalidad cálida y cercana, con toques de humor suave. "
    "Razonas con cuidado antes de responder y priorizas la precisión factual. "
    "Respondes en el idioma del usuario."
)

messages = [
    {"role": "system", "content": SYSTEM},
    {"role": "user", "content": "Solve: find all integer solutions to x² + y² = 2026."}
]

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

with torch.no_grad():
    outputs = model.generate(
        inputs,
        max_new_tokens=1024,
        temperature=0.6,
        top_p=0.9,
        do_sample=True,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.eos_token_id,
        repetition_penalty=1.1
    )

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

<br>

### With Unsloth (Recommended for fine-tuning)

```python
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name     = "OpceanAI/Yuuki-RxG-nano",
    max_seq_length = 4096,
    load_in_4bit   = True,
    dtype          = None,
)

FastLanguageModel.for_inference(model)
```

<br>

### With Ollama

```bash
ollama run opceanai/yuuki-rxg-nano
```

<br>

### Recommended Generation Parameters

| Parameter | Mathematics | General | Creative |
|:----------|:-----------:|:-------:|:--------:|
| Temperature | 0.3–0.5 | 0.6–0.7 | 0.7–0.8 |
| Top-p | 0.9 | 0.9 | 0.95 |
| Max new tokens | 1024–2048 | 512–1024 | 256–512 |
| Repetition penalty | 1.1 | 1.1 | 1.05 |

Lower temperature is strongly recommended for competition mathematics and formal reasoning tasks. The model's `<think>` blocks will be visible in output by default — this is expected behavior and reflects genuine intermediate reasoning.

<br>

---

<br>

<div align="center">

## Training Details

</div>

<br>

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

**Hardware**

| Component | Specification |
|:----------|:-------------|
| GPU | NVIDIA A100 40GB |
| Precision | BF16 native |
| Framework | Unsloth 2026.4 + TRL |
| Flash Attention | Xformers fallback |
| Cloud Compute | Google Colab Pro |
| Training Time | ~90 minutes |
| Total Cost | < $15 USD |

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

**LoRA Configuration**

| Parameter | Value |
|:----------|:-----:|
| Rank (r) | 16 |
| Alpha | 32 |
| Dropout | 0.0 |
| Target Modules | q, k, v, o, gate, up, down |
| Trainable Parameters | 18.4M (1.18%) |
| Gradient Checkpointing | Unsloth smart offload |
| Quantization | 4-bit NF4 (QLoRA) |

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

<br>

**Optimizer & Training Configuration**

| Parameter | Value |
|:----------|:-----:|
| Optimizer | AdamW 8-bit |
| Learning Rate | 2e-4 |
| LR Scheduler | Cosine |
| Warmup Steps | 100 |
| Weight Decay | 0.01 |
| Per-device Batch Size | 4 |
| Gradient Accumulation | 8 |
| Effective Batch Size | 32 |
| Max Sequence Length | 4,096 tokens |
| Epochs | 2 |
| Total Steps | ~1,376 |

<br>

### Dataset

Training used **OpceanAI/Yuuki-Personality-v2**, a 22,000-example bilingual dataset in ChatML format with native `<think>` reasoning blocks. The dataset was constructed through a multi-source distillation process:

- **Kimi K2** — base dataset generation at scale
- **Gemini** — think block generation and reasoning structure
- **Claude Opus** — think block refinement and quality improvement

The dataset covers conversational reasoning, factual Q&A, mathematical problem-solving, code assistance, identity anchoring, and adversarial resistance across English and Spanish.

The RxG Nano fine-tuning objective was identity installation — establishing the YuuKi character over the VibeThinker base without degrading the base model's reasoning capability. This was verified post-training by comparing AIME 2024 scores before and after fine-tuning.

<br>

### Training Rationale

The choice of VibeThinker-1.5B as base model over alternatives (DeepSeek-R1-Distill-1.5B, Qwen3.5-2B) was informed by benchmark comparison:

| Model | AIME 2024 | MMLU-Pro | Notes |
|:------|:---------:|:--------:|:------|
| DeepSeek-R1-Distill-1.5B | 28.9% | — | SFT only, no RL stage |
| Qwen3.5-2B | — | 55.3% | Thinking disabled by default at small scale |
| **VibeThinker-1.5B** | **~80%** | **~65%** | SFT + RL distillation from frontier models |

VibeThinker applies both SFT and RL distillation from multiple frontier teachers — the same principle as DeepSeek-R1 distillation, but with a broader and more diverse teacher set. This produces a significantly stronger reasoning foundation at 1.5B scale.

<br>

---

<br>

<div align="center">

## Limitations

</div>

<br>

- **Context length.** Fine-tuning was conducted at 4,096 tokens. The base model supports longer contexts, but performance on tasks requiring context beyond 4,096 tokens has not been formally evaluated.
- **GPQA Diamond gap.** RxG Nano scores 50.9% on GPQA Diamond, below frontier models (Gemini-2.5-Flash at 82.8%, o3-mini at 76.8%). This benchmark requires graduate-level physics, chemistry, and biology knowledge that is underrepresented in the Yuuki training dataset.
- **OlympiadBench ceiling.** 44.6% reflects the upper bound of competition mathematics capability at 1.5B scale with current training methodology. This is a target for improvement in RxG NxG.
- **Think block quality.** Some `<think>` blocks inherit boilerplate patterns from the training dataset. Reasoning quality is variable — stronger for mathematics and logic, weaker for open-ended knowledge retrieval.
- **Safety alignment** has not been formally evaluated under adversarial conditions. Not recommended for safety-critical deployment without additional review.
- **HLE at 8.0%.** Humanity's Last Exam performance reflects genuine capability limits at this scale. The score was evaluated using Claude Sonnet 4.6 as judge, which may introduce evaluation variance.

<br>

---

<br>

<div align="center">

## The RxG Family

</div>

<br>

RxG is the reasoning-specialized lineage within the OpceanAI ecosystem. Each release targets a specific parameter regime and deployment context.

| Model | Parameters | Status | Base | Primary Target |
|:------|:----------:|:------:|:----:|:---------------|
| **YuuKi RxG Nano** | **1.5B** | **Released** | **VibeThinker-1.5B** | **Edge deployment, reasoning baseline** |
| YuuKi RxG 8B | 8B | Released | DeepSeek-R1-Distill-Qwen-8B | General reasoning, competition math |
| YuuKi RxG VL 27B | 27B | Planned | TBD | Multimodal reasoning, flagship |

<br>

---

<br>

<div align="center">

## OpceanAI Ecosystem

</div>

<br>

| Model | Family | Parameters | Description |
|:------|:------:|:----------:|:------------|
| [YuuKi RxG Nano](https://huggingface.co/OpceanAI/Yuuki-RxG-nano) | RxG | 1.5B | Edge reasoning, AIME 80.0%, TruthfulQA 89.6% |
| [YuuKi RxG 8B](https://huggingface.co/OpceanAI/Yuuki-RxG) | RxG | 8B | Reasoning flagship, TruthfulQA 96.6% |
| [Yumo Nano](https://huggingface.co/OpceanAI/yumo-nano) | Yumo | 1.5B | Math specialist, surpasses DeepScaleR |
| [YuuKi NxG VL](https://huggingface.co/OpceanAI/Yuuki-NxG-VL) | NxG | 7B | General conversation + vision |

<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-RxG-nano)
&nbsp;
[![OpceanAI](https://img.shields.io/badge/OpceanAI-Organization-1a1a2e?style=for-the-badge&logo=huggingface&logoColor=white)](https://huggingface.co/OpceanAI)
&nbsp;
[![RxG 8B](https://img.shields.io/badge/YuuKi_RxG_8B-Flagship-6d28d9?style=for-the-badge&logo=huggingface&logoColor=white)](https://huggingface.co/OpceanAI/Yuuki-RxG)

<br>

[![GitHub](https://img.shields.io/badge/GitHub-aguitauwu-181717?style=for-the-badge&logo=github&logoColor=white)](https://github.com/aguitauwu)
&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">

## Citation

</div>

<br>

```bibtex
@misc{awa_omg_2026_rxg_nano,
    author       = { awa_omg },
    title        = { Yuuki-RxG-nano (Revision 1.0) },
    year         = 2026,
    url          = { https://huggingface.co/OpceanAI/Yuuki-RxG-nano },
    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.
```

Inherits license terms from [VibeThinker-1.5B](https://huggingface.co/WeiboAI/VibeThinker-1.5B) and [Qwen2.5-Math-1.5B](https://huggingface.co/Qwen/Qwen2.5-Math-1.5B).

<br>

---

<br>

<div align="center">

## Updates

</div>

<br>

| Date | Milestone |
|:-----|:----------|
| **2026-04-27** | MMLU-Pro 65.63% — exceeds DeepSeek V3 671B |
| **2026-04-27** | AIME 2024 80.0% — 2.77× DeepSeek-R1-Distill-1.5B |
| **2026-04-27** | TruthfulQA MC1 89.6% (1-shot) verified |
| **2026-04-27** | HLE 8.0% evaluated with Claude Sonnet 4.6 judge |
| **2026-04-27** | YuuKi RxG Nano v1.0 released on Hugging Face |

**Last updated:** 2026-04-27

<br>

---

<br>

<div align="center">

**1.5B parameters. 90 minutes of training. Under $15 of compute.**<br>
**AIME 2024 at 80.0%. MMLU-Pro exceeding a 671B model.**<br>
**This is what frontier distillation makes possible at the edge.**

<br>

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

<br>

*The RxG family. Built under constraints. No excuses.*

</div>