--- 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 --- # ๐ŸŒธ Yuuki โ€” Code Generation Model Trained on a Phone > **A multilingual code generation model trained entirely on a smartphone by a single person.** --- ## โš ๏ธ Disclaimer This is the **best Yuuki model available at this moment**. The latest release will be **Yuuki v0.1** โ€” once that version is published, plans for **v0.2** will begin. **Important notes:** - ๐Ÿ“ฑ This model is being trained **entirely on a smartphone** by a **single person** - ๐Ÿ“„ A **research paper** will be published soon exploring whether it's possible to train a code generation model on a mobile device - ๐Ÿšง This is an **early-stage research project**, not a production-ready model --- ## ๐ŸŒฑ 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: - โœ… The training pipeline is **functional** - โœ… The dataset is being **correctly learned** - โœ… The model is capable of generating **real, structured code-like outputs** - โœ… Early language specialization (due to dataset order) is **clearly observable** This is not a polished or production-ready model โ€” 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**. --- ## ๐Ÿ“Š Comparative Evaluation โ€” 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 โ€“ 2.23 | 1.69 โ€“ 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 | --- ## ๐Ÿงช Language Evaluation Scores ### ๐Ÿ“ Checkpoint 1400 | Language | Score | |----------|-------| | Agda | 20 | | C | 8 | | Assembly | 2 | | **Average** | **~10 / 100** | ### ๐Ÿ“ Checkpoint 2000 | Language | Score | |----------|-------| | Agda | 55 | | C | 20 | | Assembly | 15 | | **Average** | **24.6 / 100** | > ๐Ÿ“ˆ **+146% average improvement** despite minimal additional training (โ‰ˆ1.6% more steps). --- ## ๐Ÿง  Code Quality Evolution ### ๐Ÿ”ป Checkpoint 1400 โ€” 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) ``` โŒ Severe corruption, mixed tokens, runtime noise. --- #### [Test 2/5] C **Prompt:** `int main() {` ```c int main() { prostitution of '/*' (prostitution of '*' prost = $0) ``` โŒ No syntactic coherence, random token associations. --- #### [Test 3/5] Assembly **Prompt:** `mov eax,` ```asm mov eax, the International Press S. ``` โŒ No instruction continuation. --- #### [Test 4/5] Generic **Prompt:** `function test() {` ```javascript function test() { all test(1) { all a,b : a ``` โŒ 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 ``` โŒ Python not learned, heavy cross-language contamination. --- ### ๐Ÿ”บ Checkpoint 2000 โ€” 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 ``` โœ… 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 ``` โš ๏ธ 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" ``` ๐ŸŒฑ 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 ``` โš ๏ธ 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! ``` โŒ Python not yet learned (expected due to alphabetical dataset order). --- ## ๐Ÿง  Interpretation These outputs confirm that Yuuki: - ๐Ÿ“š **Learns real libraries** and language-specific tokens - ๐Ÿ—๏ธ **Shows early structure** before correctness - ๐Ÿ“Š **Reflects dataset ordering effects** honestly - ๐Ÿ“ˆ **Improves gradually**, not magically This behavior is **expected and healthy** at ~5% total training. --- ## ๐Ÿง  Key Takeaway Between **3.7% โ†’ 5.3%** training progress, Yuuki shows: - โœ… Major qualitative gains - โœ… Clear specialization trends - โœ… Rapid early learning despite CPU-only constraints This validates the project's core claim: > **Progress is real, measurable, and reproducible โ€” even at $0 cost.** --- ## ๐Ÿ“œ 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 ``` --- ## ๐Ÿ”— Links - ๐Ÿค— [Hugging Face Model](https://huggingface.co/OpceanAI/Yuuki-the-best-model) - ๐Ÿ“„ Research Paper (Coming Soon) - [Training code](https://github.com/YuuKi-OS/yuuki-training) ---

Built with patience, a phone, and zero budget.
๐ŸŒธ Yuuki Project