Yuuki-best/README.md

---
license: apache-2.0
datasets:
- bigcode/the-stack
- OpceanAI/Yuuki-dataset
language:
- en
- es
base_model:
- openai-community/gpt2
pipeline_tag: text-generation
library_name: pytorch
tags:
- code
- transformers
metrics:
- perplexity
- code_eval
new_version: OpceanAI/Yuuki-v0.1
inference: true
widget:
- text: "def fibonacci(n):"
  example_title: "Python Function"
- text: "module Main where"
  example_title: "Agda Module"
- text: "int main() {"
  example_title: "C Program"
  
---
<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;
open import Data.Nat

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 Function
--
open import Data.Nat
open import Function

open import Data.Nat
open import Data.Unit
open import Data.Nat
open import Data.Nat.Dec
open import Data.Properties.Nat
-- Dec
open import Data.Nat
open import Data.Properties
open import Data.Base
open import Data.Nat.Properties
open import Data.Unary

open import data
```

**Strong familiarity with real Agda modules and imports**

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

**C Test** ⚠️

```c
int main() {
        no-print "no"
        no-print "no"
        print "no-print "
        print "no"
        print "no"
        print "no"
        fault-print "no"
}
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, "v1"
ov 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() {
    for (i=0;i=0; i++, 1) {
   (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
- 🦙 .[ollama](https://ollama.com/aguitachan3/yuuki-best)

<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>