mindtype-tagger-v1.5

MindType's own grammar and spelling correction model, trained from scratch by Mindscend for an on-device keyboard (iOS and Android). Text never leaves the phone.

• 26.2M parameters, encoder-only transformer (8 layers, d=384), own 16k byte-level BPE tokenizer.
• Predicts one edit label per word (KEEP / DELETE / REPLACE_x / APPEND_x / case) in a single pass, so it is fast, tiny, and cannot invent words or rewrite correct text.
• Core ML package (8-bit weights) for iOS 18+, and ONNX (int8) for Android and servers.

What changed since v1.2: training pairs now include real misspellings, dropped words and tense errors (12M pairs), trained for a fixed 30,000 steps (longer training made earlier versions over-eager). Compared with v1.2: about a third fewer changes to correct real-world messages (1.9% vs 2.8%), far better at dropped words (F0.5 0.70 vs 0.08) and misspellings (0.96 vs 0.88), at the same overall F0.5.

Results (MindType evaluation suite, 4,356 held-out cases)

Settings: KEEP bias 2.0 and the plausibility guard used in the MindType apps.

Set	n	F0.5	Exact	False-positive rate	Lost protected spans
casual_clean	118	–	0.966	0.034	0.000
casual_noisy	109	0.933	0.798	0.000	0.000
coedit	579	0.464	0.109	0.000	0.000
formal_clean	1000	–	0.966	0.034	0.000
formal_noisy	1000	0.874	0.666	0.000	0.000
golden	50	0.920	0.840	0.077	0.000
preserve	1000	0.839	0.662	0.000	0.000
preserve_clean	500	–	1.000	0.000	0.000
overall	4356	0.799	0.712	0.024	0.000

False-positive rate on 2,000 held-out casual messages (never used for training; used only to choose between two training seeds): 1.9% (v1.2: 2.8%).

On the *_clean sets there is nothing to fix; the false-positive rate is what counts there.

Compared with fine-tuned small LLMs and earlier versions (same suite)

Model	Size	F0.5 ↑	Exact ↑	False positives ↓	Lost names/numbers ↓
Gemma 3 270M instruct, zero-shot	270M	0.122	0.2%	99.6%	3.9%
Gemma 3 270M + LoRA r32	270M	0.571	57.4%	23.1%	2.9%
SmolLM2 360M + LoRA r32	360M	0.663	63.1%	20.1%	0.3%
mindtype-tagger-v1.1 (keep bias 2, guard)	26M	0.748	68.3%	8.65%	0.0%
mindtype-tagger-v1.2 (keep bias 2, guard)	26M	0.800	72.4%	3.1%	0.0%
mindtype-tagger-v1.5 (keep bias 2, guard)	26M	0.799	71.2%	2.4%	0.0%

Error types added in v1.5 (separate test sets, not used for training)

Error type	n	v1.2 F0.5	v1.5 F0.5	v1.5 exact
Real misspellings ("recieve")	500	0.882	0.963	93.4%
Dropped words ("I going home")	500	0.078	0.698	39.2%
Present tense about the past ("Yesterday I go")	248	0.049	0.119	2.4%

Tense is still the weakest; it's next.

Files

File	What it is
model.safetensors	PyTorch weights (fp32): encoder + per-token classification head
config.json	architecture (layers, width, heads, vocabulary and label counts)
tokenizer.json	16k byte-level BPE tokenizer (Hugging Face tokenizers format)
labels.json	the 15,000 edit labels, index → label
plausible_pairs.json	the word pairs the plausibility guard accepts
coreml/MindTypeTagger.mlpackage	Core ML model (8-bit weights, 26 MB), inputs padded to 16/32/64/96/160 tokens; 99.96% label agreement with fp32
onnx/mindtype_tagger.onnx	ONNX model (opset 17, int8 dynamic quantization, 27 MB), inputs ids, mask (int64, [1, T]), output logits; 99.1% label agreement with fp32
metrics.json, log.jsonl, pretrain-log.jsonl, run.json	evaluation results and training logs

Training

1. Pretraining: masked language modeling on FineWeb-Edu (ODC-BY): 1 hour on ~20M sentences, then 6.8 hours on 61.7M sentences.
2. Correction: 12.02M (noisy, clean) pairs (30,000 steps at batch 256): synthetic phone-typing errors, CoEdIT GEC (Apache-2.0), and casual messages written by a local open-weight model (Qwen2.5-14B-Instruct) and filtered for correctness, plus real misspellings, dropped words and tense errors. 30% are already-correct text, so the model learns to leave good text alone.

Hardware: one RTX 3080 Ti laptop GPU, about 9 hours in total.

How it works

Text is split into words and punctuation; each word gets a label from labels.json: $KEEP, $DELETE, $REPLACE_<word>, $APPEND_<word> (insert after), or a case change. The corrected sentence is the labels applied to the words, and the model is run again (up to 3 passes) on its own output. One sentence takes about 4 ms on a phone CPU (Core ML) and on a laptop CPU (ONNX Runtime).

Inference settings used for the results: KEEP bias 2.0 (added to the $KEEP logit; higher = fewer, surer fixes). Names, numbers, emails and URLs are never changed.

Limitations

English only, sentence-level, experimental. Fixes what's there; it doesn't reword or change tone. It still rarely fixes tense. It can still change correct text (2.4% of correct sentences in our suite, 1.9% of real casual messages).

License

Apache-2.0 (weights, tokenizer, Core ML and ONNX models). Training data: FineWeb-Edu (ODC-BY), CoEdIT (Apache-2.0).