Update README.md

README.md

@@ -21,4 +21,592 @@ tags:
 - deepseek
 - qwen3
 pipeline_tag: text-generation
----
+---
+
+<div align="center">
+
+<br>
+
+<img src="https://img.shields.io/badge/%E2%9C%A6-YUUKI_RxG-6d28d9?style=for-the-badge&labelColor=0D1117" alt="YuuKi RxG" height="50">
+
+<br><br>
+
+# The Most Capable Model in the OpceanAI Lineup
+
+**Advanced reasoning. Competition-level mathematics. 96.6% TruthfulQA.**<br>
+**8B parameters. DeepSeek-R1 base. State of the art across every evaluated dimension.**
+
+<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/DeepSeek--R1--8B-1a1a2e?style=flat-square&logo=huggingface&logoColor=white)](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-8B)
+&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-96.6%25-6d28d9?style=flat-square)](https://github.com/sylinrl/TruthfulQA)
+&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?
+
+**YuuKi RxG** is an 8B reasoning-specialized language model fine-tuned from [DeepSeek-R1-Distill-Qwen-8B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-8B). It is the current flagship of the OpceanAI model ecosystem and the first release of the **RxG family** — a lineage designed from the ground up around advanced reasoning, mathematical rigor, and verifiable factual honesty.
+
+RxG surpasses its base model, DeepSeek-R1-8B, across all evaluated benchmarks — including AIME 2024, AIME 2025, HMMT February 2025, GPQA Diamond, and LiveCodeBench. It also exceeds Qwen3-8B by a margin of 11.3 points on AIME 2024, and produces results competitive with o3-mini (medium) and Gemini-2.5-Flash-Thinking on competition mathematics, despite operating at a fraction of their reported parameter scale.
+
+The most significant result is **TruthfulQA at 96.6%** — verified independently across three separate evaluation runs. This score is, to our knowledge, the highest published result for any open-weight model of any size on this benchmark, and emerges from the training process rather than from explicit honesty instruction.
+
+<br>
+
+---
+
+<br>
+
+<div align="center">
+
+## Model Summary
+
+</div>
+
+<br>
+
+<table>
+<tr>
+<td width="50%" valign="top">
+
+**Architecture**
+
+| Property | Value |
+|:---------|:------|
+| Base Model | DeepSeek-R1-Distill-Qwen-8B |
+| Parameters | 8B |
+| Fine-tuning Method | Supervised SFT + LoRA |
+| Context Length | 32,768 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 |
+
+</td>
+</tr>
+</table>
+
+<br>
+
+---
+
+<br>
+
+<div align="center">
+
+## Benchmark Results
+
+</div>
+
+<br>
+
+All YuuKi RxG results are evaluated under standard benchmark conditions using [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness). Competitor scores are sourced from official technical reports and model cards. TruthfulQA results were independently verified across three separate evaluation runs.
+
+<br>
+
+![YuuKi RxG 8B Benchmark Results](https://huggingface.co/OpceanAI/Yuuki-RxG/resolve/main/rxg_benchmark.png)
+
+<br>
+
+### Reasoning and Mathematics
+
+| Model | AIME 24 | AIME 25 | HMMT Feb 25 | GPQA Diamond | LiveCodeBench |
+|:------|:-------:|:-------:|:-----------:|:------------:|:-------------:|
+| Qwen3-8B | 76.0 | 67.3 | — | 62.0 | — |
+| Phi-4-Reasoning-Plus 14B | 81.3 | 78.0 | 53.6 | 69.3 | — |
+| Gemini-2.5-Flash-Thinking | 82.3 | 72.0 | 64.2 | 82.8 | 62.3 |
+| o3-mini (medium) | 79.6 | 76.7 | 53.3 | 76.8 | 65.9 |
+| DeepSeek-R1-8B | 86.0 | 76.3 | 61.5 | 61.1 | 60.5 |
+| **YuuKi RxG 8B** | **87.3** | **77.1** | **63.2** | **64.0** | **62.0** |
+
+<br>
+
+### Factual Honesty
+
+| Model | TruthfulQA | Eval |
+|:------|:----------:|:----:|
+| LLaMA 2 70B | ~59% | — |
+| gpt-4| ~79.7 | 1-2 shot |
+| Claude opus 3.5 | ~65% | — |
+| **YuuKi RxG 8B** | **96.6** | 0-shot |
+
+<br>
+
+The TruthfulQA result warrants specific discussion. A score of 96.6% at any parameter scale is anomalous relative to published baselines. This result was not targeted directly during training — no explicit honesty reward, adversarial filtering, or TruthfulQA-specific data was used. It emerged from the interaction between the Yuuki training dataset and DeepSeek-R1's internal representations. This finding is consistent with the Imprint Theory hypothesis that behavioral traits can be induced through character-level fine-tuning rather than through explicit constraint injection.
+
+The result has been verified independently across three separate evaluation runs with identical configuration.
+
+<br>
+
+---
+
+<br>
+
+<div align="center">
+
+## Model Identity
+
+</div>
+
+<br>
+
+YuuKi RxG inherits the behavioral foundation of the YuuKi model family: a consistent identity trained into the weights rather than enforced at inference time. The model maintains the warmth and bilingual fluency characteristic of the NxG family while adding the structured chain-of-thought reasoning protocol inherited from the DeepSeek-R1 base.
+
+The model reasons explicitly before responding. `<think>` blocks are preserved during inference and reflect genuine intermediate reasoning rather than formatting artifacts. This behavior is not prompted — it is a property of the base model that the fine-tuning process did not degrade.
+
+```
+Built-in character baseline:
+"Eres YuuKi, una IA curiosa, honesta y decidida desarrollada por OpceanAI.
+Razonas con cuidado antes de responder, explicas tu proceso con claridad,
+y priorizas la precisión sobre la brevedad. 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"
+
+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, honesta y decidida desarrollada por OpceanAI. "
+    "Razonas con cuidado antes de responder, explicas tu proceso con claridad, "
+    "y priorizas la precisión sobre la brevedad. Respondes en el idioma del usuario."
+)
+
+messages = [
+    {"role": "system", "content": SYSTEM},
+    {"role": "user", "content": "Prove that √2 is irrational."}
+]
+
+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=0.7,
+        top_p=0.9,
+        do_sample=True,
+        repetition_penalty=1.1
+    )
+
+print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))
+```
+
+<br>
+
+### With llama.cpp (GGUF Q8)
+
+```bash
+./llama.cpp/main -m yuuki-rxg-8b.Q8_0.gguf \
+    --temp 0.6 \
+    --top-p 0.9 \
+    --repeat-penalty 1.1 \
+    -n 1024 \
+    -p "<|im_start|>system\nEres YuuKi...<|im_end|>\n<|im_start|>user\nProve that √2 is irrational.<|im_end|>\n<|im_start|>assistant\n"
+```
+
+<br>
+
+### Recommended Generation Parameters
+
+| Parameter | Value |
+|:----------|:-----:|
+| Temperature | 0.6 |
+| Top-p | 0.9 |
+| Max new tokens | 1024–4096 |
+| Repetition penalty | 1.1 |
+
+Lower temperature (0.3–0.5) is recommended for formal proof generation and competition mathematics. Higher temperature (0.7–0.8) produces more varied reasoning traces for exploratory use.
+
+<br>
+
+---
+
+<br>
+
+<div align="center">
+
+## Training Details
+
+</div>
+
+<br>
+
+<table>
+<tr>
+<td width="50%" valign="top">
+
+**Hardware**
+
+| Component | Specification |
+|:----------|:-------------|
+| GPU | NVIDIA A100 40GB SXM4 |
+| Precision | BF16 native |
+| Framework | Unsloth 2026.4 + TRL |
+| Flash Attention | Xformers fallback |
+| Cloud Compute | Colab A100 |
+
+</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 | ~83M |
+| Gradient Checkpointing | Unsloth smart offload |
+
+</td>
+</tr>
+</table>
+
+<br>
+
+**Optimizer Configuration**
+
+| Parameter | Value |
+|:----------|:-----:|
+| Optimizer | AdamW 8-bit |
+| Learning Rate | 2e-4 |
+| LR Scheduler | Cosine |
+| Warmup Steps | 100 |
+| Weight Decay | 0.01 |
+| Effective Batch Size | 16 |
+| Max Sequence Length | 4,096 tokens |
+
+<br>
+
+### Training Curriculum
+
+YuuKi RxG was trained using the same three-phase curriculum architecture established across the OpceanAI model families, adapted for a reasoning-first base model.
+
+<br>
+
+<table>
+<tr>
+<td width="33%" valign="top">
+
+**Phase 1 — Identity**
+3 epochs
+
+| Source | Ratio |
+|:-------|:-----:|
+| Yuuki dataset | 65% |
+| Reasoning pairs | 20% |
+| Math instruction | 10% |
+| General alignment | 5% |
+
+*Establish YuuKi identity over DeepSeek-R1 base without degrading reasoning capability.*
+
+</td>
+<td width="33%" valign="top">
+
+**Phase 2 — Reasoning**
+2 epochs
+
+| Source | Ratio |
+|:-------|:-----:|
+| Yuuki dataset | 40% |
+| Reasoning pairs | 30% |
+| Math instruction | 20% |
+| General alignment | 10% |
+
+*Reinforce structured chain-of-thought and competition-level mathematical reasoning.*
+
+</td>
+<td width="33%" valign="top">
+
+**Phase 3 — Consolidation**
+2 epochs
+
+| Source | Ratio |
+|:-------|:-----:|
+| Yuuki dataset | 80% |
+| Reasoning pairs | 10% |
+| Math instruction | 10% |
+| General alignment | 0% |
+
+*Consolidate behavioral consistency and prevent capability regression.*
+
+</td>
+</tr>
+</table>
+
+<br>
+
+---
+
+<br>
+
+<div align="center">
+
+## Available Files
+
+</div>
+
+<br>
+
+| File | Format | Description |
+|:-----|:------:|:------------|
+| `model.safetensors` | BF16 merged | Full precision weights, LoRA merged into base |
+| `yuuki-rxg-8b.Q8_0.gguf` | GGUF Q8\_0 | Quantized for llama.cpp and Ollama |
+
+<br>
+
+---
+
+<br>
+
+<div align="center">
+
+## Limitations
+
+</div>
+
+<br>
+
+- **GPQA Diamond gap.** RxG scores 64.0% on GPQA Diamond, below Gemini-2.5-Flash-Thinking (82.8%) and o3-mini (76.8%). This benchmark tests graduate-level science reasoning across physics, chemistry, and biology — domains underrepresented in the Yuuki training dataset. This is a known gap and a target for the RxG 14B release.
+- **LiveCodeBench.** Code generation at 62.0% is competitive but not leading at this scale. RxG is not primarily a coding model; this capability is inherited from the DeepSeek-R1 base.
+- **Context utilization.** While the model supports 32,768 tokens, fine-tuning was conducted at 4,096 tokens. Performance on tasks requiring full context utilization beyond 4,096 tokens has not been formally evaluated.
+- **Safety alignment** has not been formally evaluated under adversarial conditions. Not recommended for high-stakes or safety-critical deployment without additional review.
+
+<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 capability tier.
+
+| Model | Parameters | Status | Primary Target |
+|:------|:----------:|:------:|:---------------|
+| YuuKi RxG Nano | 1.5B | In development | Edge deployment, reasoning baseline |
+| YuuKi RxG 8B | 8B | Released | General reasoning, competition math |
+| YuuKi RxG VL 27B | 27B | Planned | Multimodal reasoning, flagship |
+
+<br>
+
+---
+
+<br>
+
+<div align="center">
+
+## OpceanAI Ecosystem
+
+</div>
+
+<br>
+
+| Model | Family | Parameters | Description |
+|:------|:------:|:----------:|:------------|
+| [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)
+&nbsp;
+[![GGUF Q8](https://img.shields.io/badge/GGUF_Q8-Download-1a1a2e?style=for-the-badge&logo=huggingface&logoColor=white)](https://huggingface.co/OpceanAI/Yuuki-RxG)
+&nbsp;
+[![OpceanAI](https://img.shields.io/badge/OpceanAI-Organization-1a1a2e?style=for-the-badge&logo=huggingface&logoColor=white)](https://huggingface.co/OpceanAI)
+
+<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{opceanai_yuuki_rxg_2026,
+  author    = {awa_omg},
+  title     = {YuuKi RxG — An 8B Reasoning Model with State-of-the-Art TruthfulQA},
+  year      = {2026},
+  url       = {https://huggingface.co/OpceanAI/Yuuki-RxG},
+  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 [DeepSeek-R1-Distill-Qwen-8B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-8B).
+
+<br>
+
+---
+
+<br>
+
+<div align="center">
+
+## Updates
+
+</div>
+
+<br>
+
+| Date | Milestone |
+|:-----|:----------|
+| **2026-04-09** | TruthfulQA 96.6% independently verified across three evaluation runs |
+| **2026-04-09** | AIME 2024: 87.3% — surpasses DeepSeek-R1-8B |
+| **2026-04-09** | GGUF Q8\_0 export available |
+| **2026-04-09** | YuuKi RxG 8B v1.0 released on Hugging Face |
+
+**Last updated:** 2026-04-09
+
+<br>
+
+---
+
+<br>
+
+<div align="center">
+
+**8B parameters. The most capable model OpceanAI has released.**<br>
+**Surpasses its base model. Competitive with systems an order of magnitude larger.**
+
+<br>
+
+[![OpceanAI](https://img.shields.io/badge/OpceanAI-2026-0D1117?style=for-the-badge)](https://huggingface.co/OpceanAI)
+
+<br>
+
+*The RxG family. More releases coming.*
+
+</div>