|
|
--- |
|
|
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> |
