OpceanAI
/

Yuuki-best

@@ -18,86 +18,233 @@ metrics:
 - code_eval
 new_version: OpceanAI/Yuuki-v0.1
 ---
-# 🌸 Yuuki — Code Generation Model Trained on a Phone
-> **A multilingual code generation model trained entirely on a smartphone by a single person.**
 ---
-## ⚠️ Disclaimer
-This is the **best Yuuki model available at this moment**. The latest release will be **Yuuki v0.1** — once that version is published, plans for **v0.2** will begin.
-**Important notes:**
-- 📱 This model is being trained **entirely on a smartphone** by a **single person**
-- 📄 A **research paper** will be published soon exploring whether it's possible to train a code generation model on a mobile device
-- 🚧 This is an **early-stage research project**, not a production-ready model
 ---
-## 🌱 Best Initial Yuuki Model (Early Snapshot)
-This version of Yuuki represents the **strongest initial model** of the Yuuki project so far.
-While still early in training, this snapshot already demonstrates that:
-- ✅ The training pipeline is **functional**
-- ✅ The dataset is being **correctly learned**
-- ✅ The model is capable of generating **real, structured code-like outputs**
-- ✅ Early language specialization (due to dataset order) is **clearly observable**
-This is not a polished or production-ready model — but it is the **best starting point** Yuuki has achieved, and a **solid foundation** for future versions.
-Below are real generation samples from the current checkpoint, shown **transparently without filtering**.
 ---
-## 📊 Comparative Evaluation — Checkpoint 1400 vs Checkpoint 2000
-| Metric | Checkpoint 1400 | Checkpoint 2000 |
-|--------|-----------------|-----------------|
-| **Training Progress** | 1,400 / 37,500 (3.7%) | 2,000 / 37,500 (5.3%) |
-| **Avg Loss** | 1.70 – 2.23 | 1.69 – 2.31 |
-| **Training Speed** | ~100 sec / step | ~86 sec / step |
-| **Model Size** | 988 MB | 988 MB |
-| **Evaluated Languages** | Agda, C, Assembly, JS, Python | Agda, C, Assembly, JS, Python |
 ---
-## 🧪 Language Evaluation Scores
-### 📍 Checkpoint 1400
 | Language | Score |
-|----------|-------|
-| Agda | 20 |
-| C | 8 |
-| Assembly | 2 |
-| **Average** | **~10 / 100** |
-### 📍 Checkpoint 2000
 | Language | Score |
-|----------|-------|
-| Agda | 55 |
-| C | 20 |
-| Assembly | 15 |
 | **Average** | **24.6 / 100** |
-> 📈 **+146% average improvement** despite minimal additional training (≈1.6% more steps).
 ---
-## 🧠 Code Quality Evolution
 ### 🔻 Checkpoint 1400 — Early, Noisy, Unstable
-#### [Test 1/5] Agda
-**Prompt:** `module Main where`
 ```agda
-BLAS : Bad memory unallocation! :   50  0x72ec90a000
 module Main where
 open import Data.PropositionalEquality using (PropositionalEquality;
@@ -106,48 +253,41 @@ open import Data.Nat
 data import Data.PropositionalEquality using (Propos;
 open-Equality -PropositionalEquality) using (Prop-} (Prop-Prop-ElAdvertisements)
 ```
-❌ Severe corruption, mixed tokens, runtime noise.
----
-#### [Test 2/5] C
-**Prompt:** `int main() {`
 ```c
 int main() {
         prostitution of '/*' (prostitution of '*'
         prost = $0)
 ```
-❌ No syntactic coherence, random token associations.
----
-#### [Test 3/5] Assembly
-**Prompt:** `mov eax,`
 ```asm
 mov eax,
                the International Press              S.
 ```
-❌ No instruction continuation.
----
-#### [Test 4/5] Generic
-**Prompt:** `function test() {`
-```javascript
-function test() {
-        all    test(1) {
-        all  a,b : a
-```
-❌ Structural confusion, repeated patterns.
----
-#### [Test 5/5] Python
-**Prompt:** `def hello():`
 ```python
 def hello():
@@ -155,14 +295,22 @@ def hello():
 open program
 open import HUnit.H.P.B.P.P-E
 ```
-❌ Python not learned, heavy cross-language contamination.
----
-### 🔺 Checkpoint 2000 — Structured, Language-Aware (Best Initial Yuuki)
-#### [Test 1/5] Agda
-**Prompt:** `module Main where`
 ```agda
 module Main where
@@ -186,12 +334,13 @@ open import Data.Unary
 open import data
 ```
-✅ Shows strong familiarity with real Agda modules and imports.
----
-#### [Test 2/5] C
-**Prompt:** `int main() {`
 ```c
 int main() {
@@ -205,12 +354,15 @@ int main() {
 }
 run run
 ```
-⚠️ Syntax is unstable, but structure and intent are present.
----
-#### [Test 3/5] Assembly
-**Prompt:** `mov eax,`
 ```asm
 mov eax, "v2.2"
@@ -219,12 +371,13 @@ ov eax, "v2.2"
 ov eax, "v2._v1"
 ov eax, "v3"
 ```
-🌱 Demonstrates token-level familiarity with assembly-like patterns.
----
-#### [Test 4/5] Generic
-**Prompt:** `function test() {`
 ```javascript
 function test() {
@@ -232,74 +385,389 @@ function test() {
    (a=0; i++) {
                                  .-1       .      A
 ```
-⚠️ Logic is broken, but syntactic scaffolding is emerging.
----
-#### [Test 5/5] Python
-**Prompt:** `def hello():`
-```python
-def hello():
-        "Hello!
-        tell-tell-it -o:Hello!
-        if not exist (
-            echo "Hello!
-```
-❌ Python not yet learned (expected due to alphabetical dataset order).
 ---
-## 🧠 Interpretation
-These outputs confirm that Yuuki:
-- 📚 **Learns real libraries** and language-specific tokens
-- 🏗️ **Shows early structure** before correctness
-- 📊 **Reflects dataset ordering effects** honestly
-- 📈 **Improves gradually**, not magically
-This behavior is **expected and healthy** at ~5% total training.
 ---
-## 🧠 Key Takeaway
 Between **3.7% → 5.3%** training progress, Yuuki shows:
-- ✅ Major qualitative gains
-- ✅ Clear specialization trends
-- ✅ Rapid early learning despite CPU-only constraints
 This validates the project's core claim:
 > **Progress is real, measurable, and reproducible — even at $0 cost.**
 ---
-## 📜 License
-This project is licensed under the **Apache 2.0 License**.
 ```
 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
-    https://huggingface.co/OpceanAI/Yuuki-the-best-model/blob/main/LICENSE
 ```
 ---
-## 🔗 Links
-- 🤗 [Hugging Face Model](https://huggingface.co/OpceanAI/Yuuki-the-best-model)
-- 📄 Research Paper (Coming Soon)
-- [Training code](https://github.com/YuuKi-OS/yuuki-training)
----
-<p align="center">
-  <i>Built with patience, a phone, and zero budget.</i><br>
-  <b>🌸 Yuuki Project</b>
-</p>

 - code_eval
 new_version: OpceanAI/Yuuki-v0.1
 ---
+<div align="center">
+<br>
+<img src="https://img.shields.io/badge/%E2%9C%A6-YUUKI--BEST-000000?style=for-the-badge&labelColor=000000" alt="Yuuki Best" height="50">
+<br><br>
+# The Best Checkpoint of the $0 Phone-Trained LLM
+**Strongest initial model trained entirely on a smartphone.**<br>
+**GPT-2 architecture. Checkpoint 2000. 146% improvement over checkpoint 1400.**
+<br>
+<a href="#features"><img src="https://img.shields.io/badge/FEATURES-000000?style=for-the-badge" alt="Features"></a>
+&nbsp;&nbsp;
+<a href="https://huggingface.co/spaces/OpceanAI/Yuuki"><img src="https://img.shields.io/badge/LIVE_DEMO-000000?style=for-the-badge" alt="Demo"></a>
+&nbsp;&nbsp;
+<a href="https://github.com/sponsors/aguitauwu"><img src="https://img.shields.io/badge/SPONSOR-000000?style=for-the-badge" alt="Sponsor"></a>
+<br><br>
+[![License](https://img.shields.io/badge/Apache_2.0-222222?style=flat-square&logo=opensourceinitiative&logoColor=white)](LICENSE)
+&nbsp;
+[![Model](https://img.shields.io/badge/GPT--2-222222?style=flat-square&logo=openai&logoColor=white)](https://huggingface.co/openai-community/gpt2)
+&nbsp;
+[![Dataset](https://img.shields.io/badge/The_Stack-222222?style=flat-square&logo=databricks&logoColor=white)](https://huggingface.co/datasets/bigcode/the-stack)
+&nbsp;
+[![PyTorch](https://img.shields.io/badge/PyTorch-222222?style=flat-square&logo=pytorch&logoColor=white)](https://pytorch.org/)
+&nbsp;
+[![Transformers](https://img.shields.io/badge/Transformers-222222?style=flat-square&logo=huggingface&logoColor=white)](https://huggingface.co/docs/transformers)
+<br>
 ---
+<br>
+</div>
+## ⚠️ Important Disclaimer
+**Yuuki-best** is the **strongest checkpoint** of the Yuuki project at checkpoint 2000 (5.3% training progress). This is an **early-stage research snapshot**, not a production-ready model.
+- 🔬 **Research project** - Exploring mobile-based LLM training
+- 📱 **Single-person effort** - Trained entirely on a smartphone
+- 📄 **Research paper coming** - Full methodology and findings
+- 🚧 **Early development** - Performance will improve significantly in v0.1
+<br>
 ---
+<br>
+<div align="center">
+## What is Yuuki-best?
+</div>
+<br>
+**Yuuki-best** represents the **best checkpoint** (step 2000) of the Yuuki code generation model — a multilingual LLM trained entirely on a **Redmi 12 smartphone** with **zero cloud budget**. This checkpoint demonstrates major qualitative improvements over earlier versions, with a **146% average score increase** and clear evidence of real language learning.
+The model is based on **GPT-2 architecture** (82M parameters) and has been trained on **The Stack** dataset with 75,000 code examples. This checkpoint shows:
+- ✅ **Functional training pipeline** - Proven to work on mobile CPU
+- ✅ **Real language learning** - Generates actual Agda imports and structures
+- ✅ **Structured code outputs** - Syntactic scaffolding emerging
+- ✅ **Measurable progress** - 146% improvement in just 1.6% more training
+<br>
 ---
+<br>
+<div align="center">
+## Features
+</div>
+<br>
+<table>
+<tr>
+<td width="50%" valign="top">
+**Best Initial Checkpoint**
+Checkpoint 2000 represents the strongest model snapshot so far, with clear improvements in code structure, language awareness, and quality scores compared to earlier checkpoints.
+<br>
+**Real Language Learning**
+Generates genuine Agda imports (Cubical, Data.Nat, Function) and shows early understanding of language-specific tokens and patterns across multiple programming languages.
+<br>
+**Transparent Evaluation**
+Unfiltered generation samples showing both successes and limitations. Honest assessment of current capabilities at ~5% training progress with clear quality metrics.
+<br>
+**146% Quality Improvement**
+Average evaluation score increased from 10/100 (checkpoint 1400) to 24.6/100 (checkpoint 2000) despite only 1.6% additional training — demonstrating rapid early learning.
+</td>
+<td width="50%" valign="top">
+**Zero-Budget Training**
+Trained on a $150 Android phone with no cloud compute, no GPU acceleration, and no data center infrastructure. Proof that AI training is accessible to everyone.
+<br>
+**Multiple Quantized Formats**
+Available in GGUF format with multiple quantization levels (Q4_0, Q4_K_M, Q5_K_M, Q8_0, F32) for efficient CPU and mobile inference.
+<br>
+**Early Language Specialization**
+Due to alphabetical dataset ordering, shows strongest performance in Agda (55/100), with C, Assembly, and other languages progressively learning.
+<br>
+**Part of Complete Ecosystem**
+Integrated with CLI tools (yuy, yuy-chat), web interfaces (Yuuki-chat, Yuuki-web), and comprehensive documentation for easy deployment.
+</td>
+</tr>
+</table>
+<br>
 ---
+<br>
+<div align="center">
+## Checkpoint Comparison
+</div>
+<br>
+### Performance Metrics
+| Metric | Checkpoint 1400 | Checkpoint 2000 | Improvement |
+|:-------|:----------------|:----------------|:------------|
+| **Training Progress** | 1,400 / 37,500 (3.7%) | 2,000 / 37,500 (5.3%) | +1.6% |
+| **Average Loss** | 1.70 – 2.23 | 1.69 – 2.31 | Similar |
+| **Training Speed** | ~100 sec/step | ~86 sec/step | **14% faster** |
+| **Model Size** | 988 MB | 988 MB | Same |
+| **Evaluated Languages** | 5 languages | 5 languages | Same |
+<br>
+### Language Evaluation Scores
+<table>
+<tr>
+<td width="50%" valign="top">
+**Checkpoint 1400**
 | Language | Score |
+|:---------|:------|
+| Agda | 20 / 100 |
+| C | 8 / 100 |
+| Assembly | 2 / 100 |
+| **Average** | **10 / 100** |
+</td>
+<td width="50%" valign="top">
+**Checkpoint 2000 (Yuuki-best)**
 | Language | Score |
+|:---------|:------|
+| Agda | 55 / 100 |
+| C | 20 / 100 |
+| Assembly | 15 / 100 |
 | **Average** | **24.6 / 100** |
+</td>
+</tr>
+</table>
+<br>
+**📈 +146% average improvement** despite minimal additional training (≈1.6% more steps).
+<br>
 ---
+<br>
+<div align="center">
+## Code Quality Evolution
+</div>
+<br>
 ### 🔻 Checkpoint 1400 — Early, Noisy, Unstable
+<table>
+<tr>
+<td width="50%" valign="top">
+**Agda Test**
 ```agda
+BLAS : Bad memory unallocation! : 50 0x72ec90a000
 module Main where
 open import Data.PropositionalEquality using (PropositionalEquality;
 data import Data.PropositionalEquality using (Propos;
 open-Equality -PropositionalEquality) using (Prop-} (Prop-Prop-ElAdvertisements)
 ```
+❌ Severe corruption, mixed tokens, runtime noise
+</td>
+<td width="50%" valign="top">
+**C Test**
 ```c
 int main() {
         prostitution of '/*' (prostitution of '*'
         prost = $0)
 ```
+❌ No syntactic coherence, random token associations
+</td>
+</tr>
+<tr>
+<td width="50%" valign="top">
+**Assembly Test**
 ```asm
 mov eax,
                the International Press              S.
 ```
+❌ No instruction continuation
+</td>
+<td width="50%" valign="top">
+**Python Test**
 ```python
 def hello():
 open program
 open import HUnit.H.P.B.P.P-E
 ```
+❌ Python not learned, heavy cross-language contamination
+</td>
+</tr>
+</table>
+<br>
+### 🔺 Checkpoint 2000 — Structured, Language-Aware (Best)
+<table>
+<tr>
+<td width="50%" valign="top">
+**Agda Test** ✅
 ```agda
 module Main where
 open import data
 ```
+**Strong familiarity with real Agda modules and imports**
+</td>
+<td width="50%" valign="top">
+**C Test** ⚠️
 ```c
 int main() {
 }
 run run
 ```
+**Syntax unstable, but structure and intent present**
+</td>
+</tr>
+<tr>
+<td width="50%" valign="top">
+**Assembly Test** 🌱
 ```asm
 mov eax, "v2.2"
 ov eax, "v2._v1"
 ov eax, "v3"
 ```
+**Token-level familiarity with assembly-like patterns**
+</td>
+<td width="50%" valign="top">
+**JavaScript Test** ⚠️
 ```javascript
 function test() {
    (a=0; i++) {
                                  .-1       .      A
 ```
+**Logic broken, but syntactic scaffolding emerging**
+</td>
+</tr>
+</table>
+<br>
 ---
+<br>
+<div align="center">
+## Interpretation
+</div>
+<br>
+These outputs confirm that Yuuki at checkpoint 2000:
+<table>
+<tr>
+<td width="50%" valign="top">
+**What It Learned**
+- 📚 **Real libraries** - Actual Agda modules (Data.Nat, Cubical, Function)
+- 🏗️ **Early structure** - Syntactic scaffolding before correctness
+- 🎯 **Language tokens** - Recognizes language-specific patterns
+- 📈 **Gradual improvement** - Natural learning progression
+</td>
+<td width="50%" valign="top">
+**What It Shows**
+- ✅ **Dataset ordering effects** - Early languages learned first
+- ✅ **Real progress** - 146% quality improvement
+- ✅ **Functional pipeline** - Training works on mobile CPU
+- ✅ **Healthy behavior** - Expected at ~5% training progress
+</td>
+</tr>
+</table>
+<br>
+**This behavior is expected and healthy at ~5% total training.**
+<br>
 ---
+<br>
+<div align="center">
+## Key Takeaway
+</div>
+<br>
 Between **3.7% → 5.3%** training progress, Yuuki shows:
+- ✅ **Major qualitative gains** - 146% improvement in evaluation scores
+- ✅ **Clear specialization trends** - Strong Agda performance emerging
+- ✅ **Rapid early learning** - Despite CPU-only constraints
 This validates the project's core claim:
 > **Progress is real, measurable, and reproducible — even at $0 cost.**
+<br>
 ---
+<br>
+<div align="center">
+## Available Formats
+</div>
+<br>
+### GGUF Quantized Models
+Optimized for CPU inference with llama.cpp and Ollama.
+| Format | Size | Use Case | Quality |
+|:-------|:-----|:---------|:--------|
+| **yuuki-best-f32.gguf** | ~328 MB | Full precision baseline | Best |
+| **yuuki-best-q8_0.gguf** | ~87 MB | High quality, smaller size | Excellent |
+| **yuuki-best-q5_k_m.gguf** | ~56 MB | Balanced quality/size | Very Good |
+| **yuuki-best-q4_k_m.gguf** | ~47 MB | Good quality, efficient | Good |
+| **yuuki-best-q4_0.gguf** | ~46 MB | Most efficient, fast | Good |
+<br>
+---
+<br>
+<div align="center">
+## Usage
+</div>
+<br>
+### With Transformers (PyTorch)
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+# Load model
+model = AutoModelForCausalLM.from_pretrained("OpceanAI/Yuuki-best")
+tokenizer = AutoTokenizer.from_pretrained("OpceanAI/Yuuki-best")
+# Generate code
+prompt = "module Main where"
+inputs = tokenizer(prompt, return_tensors="pt")
+outputs = model.generate(**inputs, max_length=100, temperature=0.7)
+code = tokenizer.decode(outputs[0], skip_special_tokens=True)
+print(code)
+```
+<br>
+### With llama.cpp (GGUF)
+```bash
+# Run inference with quantized model
+./llama.cpp/main -m yuuki-best-q4_k_m.gguf \
+    -p "module Main where" \
+    -n 50 \
+    -t 4 \
+    --temp 0.7
+```
+<br>
+### With Ollama
+```bash
+# Create Modelfile
+cat > Modelfile << EOF
+FROM ./yuuki-best-q4_k_m.gguf
+TEMPLATE """{{ .Prompt }}"""
+PARAMETER temperature 0.7
+PARAMETER top_p 0.9
+EOF
+# Import and run
+ollama create yuuki-best -f Modelfile
+ollama run yuuki-best "module Main where"
+```
+<br>
+---
+<br>
+<div align="center">
+## Training Configuration
+</div>
+<br>
+<table>
+<tr>
+<td width="50%" valign="top">
+**Hardware**
+| Component | Specification |
+|:----------|:--------------|
+| Device | Redmi 12 (Android phone) |
+| CPU | Snapdragon 685 (8-core ARM) |
+| RAM | 6 GB |
+| Storage | 128 GB |
+| Training Mode | CPU only |
+| Cost | **$0.00** |
+</td>
+<td width="50%" valign="top">
+**Model Parameters**
+| Parameter | Value |
+|:----------|:------|
+| Base Model | GPT-2 (82M parameters) |
+| Dataset | The Stack + Yuuki-dataset |
+| Checkpoint | 2000 / 37,500 steps |
+| Progress | 5.3% |
+| Training Speed | ~86 sec/step |
+| Loss Range | 1.69 – 2.31 |
+</td>
+</tr>
+</table>
+<br>
+---
+<br>
+<div align="center">
+## Philosophy
+</div>
+<br>
+> **"Progress is real, measurable, and reproducible — even at $0 cost."**
+This checkpoint proves:
+- ✅ **LLM training works on mobile** - Real learning on consumer hardware
+- ✅ **Quality improves rapidly** - 146% gain with minimal training
+- ✅ **$0 budget is viable** - No cloud compute needed
+- ✅ **Anyone can contribute** - Breaking barriers to AI development
+<br>
+---
+<br>
+<div align="center">
+## Related Projects
+</div>
+<br>
+| Project | Description |
+|:--------|:------------|
+| [Yuuki-v0.1](https://huggingface.co/OpceanAI/Yuuki-v0.1) | Latest release version (coming soon) |
+| [Yuuki-3.7](https://huggingface.co/OpceanAI/Yuuki-3.7) | Intermediate checkpoint model |
+| [yuy](https://github.com/YuuKi-OS/yuy) | CLI for downloading, managing, and running Yuuki models |
+| [yuy-chat](https://github.com/YuuKi-OS/yuy-chat) | TUI chat interface for local AI conversations |
+| [Yuuki-chat](https://github.com/YuuKi-OS/Yuuki-chat) | Web-based chat interface with research modes |
+| [Yuuki-web](https://github.com/YuuKi-OS/Yuuki-web) | Official landing page and project showcase |
+| [yuuki-training](https://github.com/YuuKi-OS/yuuki-training) | Training code and scripts |
+| [Yuuki Space](https://huggingface.co/spaces/OpceanAI/Yuuki) | Web-based 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-best)
+&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;
+[![Training Code](https://img.shields.io/badge/Training_Code-GitHub-181717?style=for-the-badge&logo=github&logoColor=white)](https://github.com/YuuKi-OS/yuuki-training)
+<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;
+[![YUY Chat](https://img.shields.io/badge/Yuy_Chat-GitHub-181717?style=for-the-badge&logo=github&logoColor=white)](https://github.com/YuuKi-OS/yuy-chat)
+&nbsp;
+[![Sponsor](https://img.shields.io/badge/Sponsor-GitHub_Sponsors-ea4aaa?style=for-the-badge&logo=githubsponsors&logoColor=white)](https://github.com/sponsors/aguitauwu)
+</div>
+<br>
+---
+<br>
+<div align="center">
+## Community
+</div>
+<br>
+Join the Yuuki community:
+- 💬 [Discord Server](https://discord.gg/j8zV2u8k) - Chat with other users and contributors
+- 🐦 [Twitter Updates](https://twitter.com/aguitauwu) - Follow development progress
+- 📺 [GitHub](https://github.com/aguitauwu) - Star repos and contribute
+- 💖 [GitHub Sponsors](https://github.com/sponsors/aguitauwu) - Support the project
+<br>
+---
+<br>
+<div align="center">
+## Acknowledgments
+</div>
+<br>
+- **HuggingFace** - Infrastructure and transformers library
+- **BigCode** - The Stack dataset
+- **The ML community** - For inspiration and support
+- **Everyone following along** - Your interest makes this worthwhile
+<br>
+---
+<br>
+<div align="center">
+## License
+</div>
+<br>
 ```
+Apache License 2.0
+Copyright (c) 2026 Yuuki Project
 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.
 ```
+**You can use Yuuki commercially, modify it, distribute it. Just give credit.** ✅
+<br>
 ---
+<br>
+<div align="center">
+**Built with patience, a phone, and zero budget.**
+<br>
+[![Yuuki Project](https://img.shields.io/badge/Yuuki_Project-2026-000000?style=for-the-badge)](https://huggingface.co/OpceanAI)
+<br>
+*The best checkpoint so far. More improvements coming soon.* 🌸
+</div>