Greek Transformer Language Model for FUTO Keyboard (~36.1M)
Custom Modern Greek (el) predictive text and autocorrect Transformer Language Model (~36.15M parameters) built for the FUTO Android Keyboard.
- Architecture:
LlamaForCausalLM(9 layers, 512 hidden size, 1376 intermediate, 8 heads, tied embeddings) - Quantization Formats:
Q6_K(~36.8 MB),Q8_0(~45.4 MB),F16(~85 MB) - Tokenizer: SentencePiece Unigram (
15,008vocab,treat_whitespace_as_suffix=True) - Control Tokens:
<XBU>,<XBC>,<XEC>for FUTO autocorrect protocol
🚀 Quick Start & Mobile Deployment
1. Requirements
To support Greek in-word typing, Unicode codepoints, and spatial keystroke decoding, install the APK built from the Greek-enabled fork: 👉 https://github.com/stamchry/android-keyboard
2. Download Model
Download el_keyboard_v3_Q6_K.gguf (or el_keyboard_Q6_K.gguf) to your Android device storage.
3. Setup in FUTO Keyboard
- Open FUTO Keyboard Settings → Help & About.
- Tap "Version code" 8 times rapidly to unlock Developer Settings.
- Open Settings → Developer → toggle "Allow transformer models on non QWERTY layouts" to ON.
- Go to Settings → Predictive Text → Transformer Models.
- Tap Actions (top-right) → Import from file → select
el_keyboard_v3_Q6_K.gguf. - Ensure it displays: "Model is set to default for el".
📊 Benchmark Results
| Metric / Benchmark | Result |
|---|---|
| Next-Word Top-1 Accuracy | 39.33% |
| Next-Word Top-3 Accuracy | 53.06% |
| Next-Word Top-10 Accuracy | 64.48% |
| Accent Restoration Accuracy | 38.60% |
| Synthetic Typo Autocorrect | 37.40% |
🔗 Repository
Source training code and pipeline: https://github.com/stamchry/greek-keyboard-lm