GGUF Generator

World's first Retrainable GGUF — train, LoRA-adapt & quantize LLMs on Android.

The only app in the world that lets you create, train, retrain, LoRA-adapt, and quantize GGUF language models entirely on your Android device.

"Retrainable GGUF" means the model you export embeds its training history and optimizer state — reload it anytime to continue learning right where you left off.

No cloud, no subscription, no data ever leaves your phone.

---

1. Overview — What Is a Retrainable GGUF?

A standard GGUF file is inference-only. GGUF Generator embeds the training corpus and Adam optimizer state inside every exported GGUF, making it possible to reload the file and resume training seamlessly.

Think of it like a textbook that remembers every lesson — always ready for the next one.

This is a world first in the llama.cpp ecosystem.

---

2. Key Features

■ Interactive Learning (New GGUF)

Teach a small LLM from scratch through chat — instruction by instruction, response by response.

- Two-step input (instruction → response): train and verify inference instantly

- Re-learn: reinforce the last pair with repeated training passes

- Response Learn: use the model's last reply as instruction to teach conversation flow

- Batch pair import (JSONL): vocab rebuild + full retrain

- Additional Training: memorize mode ON trains until all pairs are reproduced; OFF trains to target loss

- Stop Training: halt background training at any time; current trained state is preserved

- GGUF export: specify a model name; F32 / F16 / Q8_0 — ready for llama.cpp or Ollama

■ Interactive Learning (LoRA)

Apply the same chat-based teaching to a LoRA adapter on top of any existing GGUF model.

- Supported architectures: LFM2 / Llama-3 / Gemma3 / Qwen3 (under 1B recommended)

- Adapter GGUF embeds training pairs + Adam state → reload and resume instantly

- Specify model name on export; stop background training mid-run

■ Pair Edit & Generation (Auto Training Data)

Use an external LLM server to auto-generate and edit training pairs.

- Generate diverse Q&A pairs from web URLs, Wikipedia, or a free-form theme prompt

- Edit, translate, rephrase, or derive new pairs using the LLM

- Translation expansion and paraphrase amplification for multilingual datasets

■ Quantization Conversion

Convert any GGUF file to a different quantization type on-device.

- Supported: F32 / F16 / Q8_0 / Q6_K / Q5_K_M / Q5_1 / Q5_0 / Q4_K_M / Q4_1 / Q4_0 / Q3_K_M / Q2_K

■ Dark Mode

Switch between Auto / Light / Dark in Settings. Auto follows the system theme.

---

3. Who Is This For?

- Anyone who wants a personalized small assistant model trained entirely on their phone

- Developers and researchers building domain-specific LLMs (FAQ bots, specialized assistants)

- Users who want to add custom knowledge or language style to an existing model via LoRA

- Privacy-first AI users — data never leaves the device

- Students and enthusiasts learning machine learning hands-on

---

4. Getting Started

A) Build a model from scratch

1. Open "Interactive (New GGUF)" → tap "New" to initialize a model

2. Enter instruction → "New Learn" → enter response → "New Learn"

3. Tap "Export GGUF", enter a model name, and save

B) Fine-tune an existing model with LoRA

1. Open "Interactive (LoRA)" → "Load model" to load a base GGUF

2. Follow the same two-step input. Tap "Export adapter", enter a name, and save.

C) Auto-generate pairs → batch train

1. Open "Pair Edit & Gen" → generate pairs from a URL, Wikipedia, or theme

2. Save as JSONL, then go to "Interactive" → "Import pairs" for full automatic training

LLM server (required for pair generation):

Point the app at an Ollama-compatible server (same device or local network) via Settings → Server URL.