OpceanAI/Yuuki-best

🌸 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

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() {

int main() {
        prostitution of '/*' (prostitution of '*'
        prost = $0)

❌ No syntactic coherence, random token associations.

---

[Test 3/5] Assembly

Prompt: mov eax,

mov eax,

               the International Press              S.

❌ No instruction continuation.

---

[Test 4/5] Generic

Prompt: function test() {

function test() {
        all    test(1) {
        all  a,b : a

❌ Structural confusion, repeated patterns.

---

[Test 5/5] Python

Prompt: def hello():

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

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() {

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,

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() {

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():

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
• 📄 Research Paper (Coming Soon)
• Training code

---

Built with patience, a phone, and zero budget.
🌸 Yuuki Project