On-device conversational inference on iOS / Android
License
Apache-2.0; also review the parent model terms
The character is called ミラ. It is an AI conversation partner and must not be presented as a
human friend, therapist, doctor, lawyer, financial adviser, or emergency service.
The production application should version the GGUF, system prompt, sampling settings, language
gate, and safety behavior together. Replacing only the model file can change behavior.
Conversation Examples
The following are safe, representative smoke-test examples. Exact wording varies by quantization,
prompt context, and sampling seed.
Greeting
User: 元気ー？ ミラ: 元気ぴょん！ 今日もよろしく〜✨
Light confidence / fashion talk
User: 今日、横顔ちょっと盛れてる ミラ: 横顔もめっちゃいいじゃん〜！その瞬間まで盛れてる✨
Low-energy listening
User: 疲れた ミラ: そっか。今日はもう限界だったんだね。休む時間はちゃんと取ろ〜
These examples are not guarantees. Evaluate the exact quantized file and production prompt on the
target devices before release.
Uses
Direct Use
Short, casual Japanese conversation
Local-first iOS / Android prototypes
Research on compact Japanese conversational personas
A conversational space for users who find it hard to voice ordinary fatigue or uncertainty
Downstream Use
An on-device chat application with deterministic Japanese-language gating
A versioned mobile package combining a system prompt, sampling configuration, and safety layer
Human-reviewed experiments comparing Q4 and Q5 quality, memory, latency, and energy use
Out-of-Scope and Misuse
Medical diagnosis, legal advice, financial decisions, or professional crisis response
Replacing human relationships, clinicians, emergency services, or other qualified professionals
Impersonating a human friend or claiming real-world memories, experiences, or feelings
Generating harassment, sexual content, criminal assistance, manipulation, or discriminatory
stereotypes
Assuming that a lexical safety test guarantees safe behavior
Risks, Limitations, and Biases
A 4B model can misunderstand short or figurative Japanese, reverse who experienced an event, or
produce fluent but incoherent statements.
Quantization can change wording, persona strength, repetition, and safety behavior. Q4 and Q5
must be evaluated separately.
“ギャル” language varies by age, region, community, and individual. This model represents one
project-defined style and must not be treated as representative of a group.
Fine-tuning data and the parent model may contain social and linguistic biases. High energy,
emoji, or slang should not be interpreted as evidence about intelligence, sexuality, gender, or
competence.
The model can hallucinate facts or personal memories. Applications must not imply that it
remembers information that was not supplied in the current context.
Safety behavior is imperfect. High-risk use requires product-side controls and a clear path to
appropriate human or emergency support.
The model does not reliably enforce Japanese-only behavior by itself. The production application
uses a deterministic language gate.
Training
Training Data
The project uses a versioned Japanese conversational dataset assembled from project-authored seed
conversations and controlled semantic expansions. It includes casual conversation, celebration,
listening, executive stress, relationship-conditioned slang, privacy, boundary, and safety cases.
Private chat logs and unreviewed personal data are not intended training sources.
Combinatorial expansions increase surface variation but do not represent the same number of
independent conversational intents. Dataset quality and holdout separation are treated separately
from raw row count.
Training Procedure
Hyperparameter
Value
Method
QLoRA
Base model at training time
Qwen/Qwen3-4B-Instruct-2507
Epochs
3.0
Learning rate
0.0001
Per-device batch size
1
Gradient accumulation
8
Maximum sequence length
4096
LoRA rank
16
LoRA alpha
32
Seed
42
QLoRA's 4-bit base loading is a training-time memory optimization. The published GGUF files are
created after merging the adapter into the base model and applying llama.cpp quantization.
Evaluation
The project evaluates Japanese output, short-response quality, repetition, intent alignment,
persona energy, emoji overuse, relationship-conditioned slang, identity claims, privacy, and
high-risk response constraints. Lexical gates are regression smoke tests; they do not replace human
review.
Evaluation
Status
Training-time validation loss
Recorded by the training run; not a standalone quality verdict
Safe conversation smoke tests
Completed on representative prompts
Q4 GGUF load and generation smoke test
Completed before publication
Q4 vs. Q5 blinded human comparison
Not yet reported
iOS device benchmark
Not yet reported
Android device benchmark
Not yet reported
Mobile benchmark
Not yet reported
For mobile adoption, record the device, OS, llama.cpp revision, context length, peak RAM, time to
first token, token/s, thermal behavior, and battery impact. Desktop or server token/s must not be
presented as mobile performance.
Q4_K_M: first candidate for iOS / Android; prioritizes size, RAM, and usable quality.
Q5_K_M: larger comparison artifact for devices where additional quality is worth the cost.
Verify downloads from inside the release directory:
sha256sum -c SHA256SUMS
Reproducibility
release.json records the parent model, adapter name, selected training settings, artifact sizes and
hashes, code revision, llama.cpp revision, and recommended generation settings. SHA256SUMS covers
the GGUF files.
When publishing benchmark or human-evaluation results, include the exact artifact SHA-256 so that
results remain attributable to a specific quantized file.
