mydeskmate/grammar-v5

Grammar Correction v5 (MyDeskMate)

English grammar-correction engine: T5-base (223M params) fine-tuned by MyDeskMate, plus the int8 ONNX graphs the live CPU API actually serves.

Measured (14-16 Oct 2026, official scorers)

what	measured	target	verdict
JFLEG fluency (GLEU)	62.78	>= 60	PASS
Never harms good text (10,000 clean sentences)	2.39% touched	< 5%	PASS
Over-correction (2,000-sentence protocol)	3.95%	< 5%	PASS
Speed per sentence (CPU, intra_op=2)	104 ms	< 200 ms	PASS
Determinism / safety / 512-token context	pass / 0 toxic / 515->515	-	PASS
CoNLL-2014 academic exam (F0.5)	47.35	>= 65	FAIL (open)
BEA-2019 advanced exam (F0.5)	33.90	>= 70	FAIL (open)

The two FAIL rows are published as measured, never hidden. They need GPU-scale models; this checkpoint is a fast, private, CPU-first small model.

Files

• config.json, generation_config.json, tokenizer.json, tokenizer_config.json
• model.safetensors (850 MB fp32 teacher; the quantized ORT graphs below are what serve)
• v5_decoder_step_xkv_int8.onnx + v5_encoder_pc.onnx (the exact graphs measured at 104 ms)

Training data / license note

Trained on research corpora (JFLEG, CoEdit, WikiText-103 subsets). Weights are provided free for demo/education under license: other -- check each source dataset's license before any commercial use.

Live demo: https://tools.mydeskmate.in/grammar.html Served by mydeskmate/grammar-checker (Hugging Face Space).