Update README.md

README.md

@@ -11,4 +11,289 @@ pipeline_tag: text-generation
 library_name: transformers
 tags:
 - code
----
+---
+
+# 🌸 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
+
+    http://www.apache.org/licenses/LICENSE-2.0
+```
+
+---
+
+## 🔗 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>