English tap-typing model for FUTO Keyboard (unofficial fine-tune)
A drop-in English language model for FUTO Keyboard
0.1.30, fine-tuned from FUTO's stock English model (ml4_q6_k.gguf) to fix tap typos
more often.
Unofficial. Not made, reviewed or endorsed by FUTO. This is a modified version of FUTO's model (see License).
| File | futo-english-tap-v4b-q8_0.gguf (39 MB, Q8_0) |
| SHA-256 | c8943241f994d039a9ea4e0b335a92e8b7e6580bd188d20e8d327a4f9130f625 |
| Architecture | Same as stock: llama, 8 layers, dim 512, 36M parameters, FUTO tokenizer (15,008) |
| Features | Same as stock: base_v1 inverted_space xbu_char_autocorrect_v1 xc0_swipe_typing_v1 char_embed_mixing_v1 |
Install
FUTO Keyboard → Settings → Languages & Models → English → import model → pick the .gguf.
Keep the stock model to switch back.
Results
Evaluated with FUTO's own native inference code (pinned 0.1.30 source), beam 3, on held-out English text with simulated tap noise. Each case is one target word in context.
Fresh holdout (2,048 Wikipedia cases never used in training or selection, opened once):
| Stock | This model | |
|---|---|---|
| Typos corrected (applied) | 73.1% | 82.5% (+9.3 pts, 95% CI +7.8 to +11.0) |
| Typo top-1 suggestion | 80.0% | 86.1% |
| Correct words wrongly changed | 2.05% | 1.07% |
| Next-word in top 3 | 35.4% | 38.6% (+3.2 pts, CI +2.1 to +4.4) |
Earlier sealed test (2,048 Wikipedia + Tatoeba cases): typos corrected 72.7% → 82.8%, next-word top 3 31.0% → 33.7%, wrongly changed words 1.17% → 1.22% (24 → 25 cases). It formally failed one preregistered check, an absolute 99% clean-word floor that stock also misses on that set (98.83%).
Q8_0 vs full precision (2,048 dev cases): same typo corrections (1,577) and same wrongly changed words (18); 99.6% identical final outputs. Q6_K lost about 2 pts, so it is not published.
Limitations
- Typo noise is synthetic (QWERTY geometry), not recorded touches. Real-world gains may differ.
- Tested on lowercase ASCII words only. Dictionary, capitalization and the full app pipeline are outside the benchmark. One user's informal phone trial: typos felt handled better.
- Swipe is unchanged in practice: in FUTO 0.1.30, swiped words don't reach this model's correction path. The swipe feature flag is kept only to match stock.
- English only. Latency and battery were not measured; Q8_0 is about 30% larger than stock.
Training
- Start: stock weights, dequantized to full precision (verified against FUTO's native code).
- Data: English Wikipedia (
wikimedia/wikipedia, 20231101.en, first shard, CC BY-SA) and Tatoeba English sentences (CC0). No private or user data. - Tasks: next-word language modeling plus noisy-tap → correct-word examples in FUTO's input format; 3,000 steps, batch 4, learning rate 1e-5, Apple MPS.
- Selection: chosen on dev data among about 25 checkpoints; confirmed on the holdouts above.
License
FUTO Source First License 1.1-kb (LICENSE.md), inherited from FUTO's model. Non-commercial use and free distribution only. Modified by fine-tuning as described above. Training text includes Wikipedia content under CC BY-SA.
- Downloads last month
- 37
8-bit