README.md

---
license: apache-2.0
datasets:
- bigcode/the-stack
language:
- en
- es
base_model:
- openai-community/gpt2
pipeline_tag: text-generation
library_name: pytorch
tags:
- code
- transformers
metrics:
- perplexity
---

# 🌸 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;
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():

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

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
```
✅ Shows strong familiarity with real Agda modules and imports.

---

#### [Test 2/5] C
**Prompt:** `int main() {`

```c
int main() {
        no-print "no"
        no-print "no"
        print "no-print "
        print "no"
        print "no"
        print "no"
        fault-print "no"
}
run run
```
⚠️ Syntax is unstable, but structure and intent are present.

---

#### [Test 3/5] Assembly
**Prompt:** `mov eax,`

```asm
mov eax, "v2.2"
ov eax, "v1"
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() {
    for (i=0;i=0; i++, 1) {
   (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>