yesterday

ollama run aguitachan3/yuuki-3.7:q4_0

Details

yesterday

347a9ecf170b · 61MB ·

gpt2
·
81.9M
·
Q4_0
You are Yuuki 3.7, a code generation model trained on a smartphone.
Apache 2.0
{ "stop": [ "</s>" ], "temperature": 0.7, "top_p": 0.9 }
{{ .Prompt }}

Readme

28285.gif

Yuuki-3.7

Code generation model trained entirely on a smartphone with zero budget. 82M parameters. GPT-2 architecture. Checkpoint 1400 of 37,500 steps.

HuggingFace Sponsors License

Quick Start

ollama run aguitachan3/yuuki-3.7

About

Yuuki-3.7 represents an intermediate checkpoint of the Yuuki project at step 1400, capturing the model during early learning phases at 3.7% of total planned training. This checkpoint serves as a valuable baseline for researchers interested in observing training progression and early-stage model behavior.

Training occurred entirely on a Redmi 12 smartphone with Snapdragon 685 processor in CPU-only mode, demonstrating that meaningful model development remains possible without specialized hardware infrastructure. The checkpoint preserves model state during a critical phase where fundamental language patterns begin emerging across the training dataset.

Performance metrics reflect early-stage learning with lower absolute scores compared to later checkpoints, though the progression from this baseline demonstrates rapid improvement rates. The model exhibits beginning understanding of code structure in Agda and C, with minimal exposure to languages appearing later in the alphabetically-ordered training data.

Supported languages: Agda (early stage), C (initial learning), Assembly (minimal), Python (not yet reached)

Input: Text prompts for code generation

Output: Early-stage code generation across limited languages

Training device: Redmi 12 smartphone (Snapdragon 685, 6GB RAM)

Training cost: $0.00

Models

Available Quantizations

ollama run aguitachan3/yuuki-3.7:q4_k_m  # Recommended
Quantization Size Description
q4_0 46 MB Most efficient format with fastest inference speed
q4_k_m 47 MB Recommended format balancing quality and performance
q5_k_m 56 MB Higher quality with moderate size increase
q8_0 87 MB High quality format for quality-sensitive applications
f32 328 MB Full precision format for research and baseline comparisons

The q4_k_m quantization provides optimal balance between model quality and inference efficiency for most research use cases.

Usage

Interactive Generation

The model responds to code generation prompts with early-stage outputs reflecting its position in training progression. Example interactions demonstrate emerging language understanding:

ollama run aguitachan3/yuuki-3.7
>>> Generate a fibonacci function in Python
>>> Create a simple linked list in C
>>> Write an Agda module for natural numbers

API Integration

Researchers can interact with the checkpoint through standard Ollama API patterns for training progression analysis:

import ollama

response = ollama.generate(
    model='aguitachan3/yuuki-3.7',
    prompt='def factorial(n):'
)

print(response['response'])

Custom Configuration

Checkpoint behavior can be adjusted through parameter configuration for experimental purposes:

ollama run aguitachan3/yuuki-3.7 \
    --temperature 0.7 \
    --num-predict 200 \
    --top-p 0.9

Modelfile Customization

Research configurations can define specific system prompts acknowledging the checkpoint’s early training stage:

FROM aguitachan3/yuuki-3.7:q4_k_m

PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER num_ctx 2048
PARAMETER stop "</s>"

SYSTEM """You are Yuuki 3.7, an intermediate checkpoint of a code generation model trained on a smartphone.

You are in early training stages learning multiple programming languages. Generate code following basic language conventions."""

Performance

The checkpoint demonstrates early-stage learning with quality scores reflecting its position at 3.7% of total training. Performance varies significantly by language based on alphabetical dataset exposure:

Language Quality Score Training Status
Agda 20 / 100 Early learning with basic pattern recognition
C 8 / 100 Initial exposure with minimal structure
Assembly 2 / 100 Negligible exposure and understanding
Python 0 / 100 Not yet encountered in training sequence

Comparison with checkpoint 2000 reveals 146% average improvement across languages despite only 1.6% additional training, demonstrating rapid learning dynamics characteristic of early training phases.

Technical Details

The architecture follows GPT-2 specifications with 82 million parameters organized across 12 transformer layers. Each layer contains 12 attention heads operating on 768-dimensional embeddings. The vocabulary encompasses 50,257 tokens with context window capability extending to 1,024 tokens.

Training execution on mobile hardware proceeds at approximately 100 seconds per training step using the Snapdragon 685 processor’s 8 ARM cores. Cumulative training time for this checkpoint exceeds 38 hours of continuous processing. Subsequent checkpoints demonstrate improved training efficiency as optimization routines stabilize.

Dataset processing follows alphabetical ordering by programming language, creating intentional exposure bias toward languages like Agda and C. This ordering strategy produces measurable effects on model performance, with early-alphabet languages showing stronger pattern recognition compared to languages encountered later in training sequences.

Capabilities and Limitations

The checkpoint exhibits emerging understanding of Agda syntax with basic library import patterns. C language capabilities show initial structure recognition though semantic correctness remains inconsistent. Assembly code generation demonstrates minimal familiarity with instruction formats. The model operates efficiently on CPU hardware making it suitable for edge deployment research scenarios.

Limitations reflect the checkpoint’s position at 3.7% of planned training. Language coverage remains severely constrained by alphabetical dataset ordering. Quality metrics indicate early research-stage outputs unsuitable for production applications. The checkpoint serves primarily as a training progression baseline rather than a functional code generation tool. Continued training through subsequent checkpoints addresses these limitations progressively.

Checkpoint Comparison

Comparative analysis with checkpoint 2000 quantifies training progression dynamics:

Metric Checkpoint 1400 Checkpoint 2000 Change
Training Progress 3.7% 5.3% +1.6%
Agda Score 20 / 100 55 / 100 +175%
C Score 8 / 100 20 / 100 +150%
Assembly Score 2 / 100 15 / 100 +650%
Training Speed ~100 sec/step ~86 sec/step +14% faster
Average Quality 10 / 100 24.6 / 100 +146%

These metrics demonstrate rapid quality improvements during early training phases, with particularly dramatic gains in languages with minimal initial exposure like Assembly.

Integration

VS Code integration through Continue extension supports checkpoint analysis by adding configuration to .continue/config.json:

{
  "models": [{
    "title": "Yuuki 3.7",
    "provider": "ollama",
    "model": "aguitachan3/yuuki-3.7"
  }]
}

Cursor IDE configuration accepts the model through Settings under Models with identifier ollama/aguitachan3/yuuki-3.7.

Open WebUI provides interface access for checkpoint interaction and comparison studies. Standard container deployment with volume mounting enables persistent configuration storage.

Research Value

This checkpoint provides baseline measurements for studying mobile-based model training progression. The early training stage captures fundamental pattern formation before higher-order language understanding emerges. Researchers can analyze quality score trajectories, training dynamics, and hardware efficiency metrics using this checkpoint as reference.

The checkpoint demonstrates that even partial training on consumer hardware produces measurable language understanding. Performance improvements between checkpoints validate training methodology and dataset curation strategies. Comparison studies enable analysis of learning rate dynamics and language acquisition patterns under resource-constrained training conditions.

Philosophy

The Yuuki project demonstrates that AI research remains accessible through consumer hardware and sustained effort. Training checkpoints document the progression from random initialization through emerging language understanding, validating that expensive infrastructure represents convenience rather than necessity for machine learning development.

The model name combines Yuki (snow in Japanese) with Yuu from Girls’ Last Tour, forming Yuuki meaning courage. This naming captures the project’s spirit of attempting unconventional approaches despite limited resources.

Related Resources

The primary checkpoint yuuki-best (step 2000) demonstrates significantly improved performance for production evaluation. The complete v0.1 release will represent full training completion across all 37,500 steps. Web applications including Yuuki Chat and Yuuki Web provide alternative interaction methods. Command-line tools yuy and yuy-chat support local model management and terminal interaction. Community resources include Discord, Reddit, and GitHub organization repositories.

License

Apache License 2.0

Copyright (c) 2026 Yuuki Project

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.

See http://www.apache.org/licenses/LICENSE-2.0

Reference