Update README.md

README.md

@@ -19,6 +19,633 @@ tags:
 - fine-tuned
 - chat
 - deepseek
-- qwen3
+- 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_v5.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>