TonyRaju/visheratin-t5-tiny-grammar-correction-onnx-int8

visheratin/t5-efficient-tiny-grammar-correction — ONNX INT8 Quantized

This is an ONNX INT8 dynamically quantized version of visheratin/t5-efficient-tiny-grammar-correction for grammatical error correction in English.

This is a derivative work. All credit for the original model goes to Alexander Visheratin (@visheratin). This repository only provides a quantized ONNX conversion for easier deployment. We do not claim ownership of the model architecture or weights.

Original Model

Original Repository	visheratin/t5-efficient-tiny-grammar-correction
Original Author	Alexander Visheratin (@visheratin)
Architecture	T5-efficient-tiny (encoder-decoder)
Parameters	15.6M
Training Data	C4_200M subset + nlpaug typo augmentation
License	No license specified in original repo

What’s in this repo?

• encoder_model.onnx — Encoder graph (INT8 quantized)
• decoder_model.onnx — Decoder graph (INT8 quantized)
• decoder_with_past_model.onnx — Decoder with KV-cache (INT8 quantized)
• config.json, generation_config.json — Model configuration
• tokenizer.json, tokenizer_config.json, special_tokens_map.json — Tokenizer files

Quantization Details

Step	Detail
Export	PyTorch → ONNX FP32 via HuggingFace Optimum (ORTModelForSeq2SeqLM)
Quantization	onnxruntime.quantization.quantize_dynamic with QuantType.QInt8
Validation	Verified ONNX FP32 output matches PyTorch FP32 exactly (0 delta) before quantizing

Benchmark Results — This Model

Benchmarked on a 50-sentence English grammar correction test set (2-thread CPU, Google Colab).

Quality

Format	BLEU	chrF++	Exact Match	Δ BLEU	Δ chrF++
PyTorch FP32 (baseline)	66.08	82.69	48%	—	—
ONNX INT8 (this repo)	64.51	82.28	44%	-1.57	-0.42

Speed & Size

Metric	PyTorch FP32	ONNX INT8 (this repo)
P50 Latency	85 ms	60 ms
P95 Latency	112 ms	107 ms
Disk Size	59 MB	55 MB
Throughput	—	60.33 sentences/sec

Quantization Stability

Metric	Value
Sentences with changed output (INT8 vs FP32)	2/50 sentences changed
Deployment Scorecard Verdict	GO

Smallest and fastest model. Ideal for mobile/edge deployment where size is critical. Statistically identical quality to visheratin-mini (p=0.824).

---

Cross-Model Comparison (7 Models Benchmarked)

This model was benchmarked alongside 6 other grammar correction models. All models were evaluated on the same 50-sentence test set under identical conditions.

PyTorch FP32 Baseline (all models)

Model	Params	BLEU	chrF++	P50 (ms)	Size (MB)	Exact Match
prithivida	222.9M	76.88	86.40	602	850	58%
coedit-small	60.5M	77.42	86.84	226	231	60%
vennify-t5-base	222.9M	78.57	87.87	539	850	58%
aventiq-t5-small	60.5M	34.43	45.88	213	231	16%
visheratin-mini	31.2M	66.84	81.77	133	119	46%
visheratin-tiny ←	15.6M	66.08	82.69	85	59	48%
pszemraj-small	77.0M	57.11	77.13	363	294	28%

ONNX INT8 Quantized (all models)

Model	Params	BLEU (FP32)	BLEU (INT8)	chrF++ (INT8)	P50 (ms)	Disk (MB)	Exact Match	Scorecard
prithivida	222.9M	76.88	74.44	84.79	222	426	54%	NO-GO
coedit-small	60.5M	77.42	77.60	86.69	117	152	60%	GO
vennify-t5-base	222.9M	78.57	77.24	86.74	224	426	56%	GO
aventiq-t5-small	60.5M	34.43	40.66	51.95	120	152	22%	GO*
visheratin-mini	31.2M	66.84	66.65	82.67	78	93	46%	GO
visheratin-tiny ← this model	15.6M	66.08	64.51	82.28	60	55	44%	GO
pszemraj-small	77.0M	57.11	3.15	18.72	138	153	0%	NO-GO

Statistical Significance (Bootstrap, 1000 resamples)

Key cross-model comparisons (INT8 variants):

Model A	Model B	chrF++ A	chrF++ B	p-value	Significant?
coedit-small	vennify-t5-base	86.69	86.74	1.000	No (identical)
coedit-small	visheratin-mini	86.69	82.67	0.038	Yes
coedit-small	visheratin-tiny	86.69	82.28	0.010	Yes
coedit-small	prithivida	86.69	84.79	0.324	No
visheratin-mini	visheratin-tiny	82.67	82.28	0.824	No (identical)
prithivida	vennify-t5-base	84.79	86.74	0.324	No

Recommendation

Use Case	Recommended Model	Why
Best overall (desktop)	coedit-small INT8 (152 MB)	Highest INT8 quality (chrF++ 86.69), 117ms latency, only 1 regression
Smallest (mobile/edge)	visheratin-tiny INT8 (55 MB)	60ms latency, 55 MB disk, acceptable quality (chrF++ 82.28)
Highest baseline quality	vennify-t5-base INT8 (426 MB)	chrF++ 86.74, but 3x larger than coedit-small for no significant quality gain

---

Usage

from optimum.onnxruntime import ORTModelForSeq2SeqLM
from transformers import AutoTokenizer

model_id = "YOUR_USERNAME/visheratin-t5-tiny-grammar-correction-onnx-int8"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = ORTModelForSeq2SeqLM.from_pretrained(model_id)

text = "(none — empty string)She go to school yesterday"
inputs = tokenizer([text], return_tensors="pt", max_length=128, truncation=True)
outputs = model.generate(**inputs, max_new_tokens=128, num_beams=1, repetition_penalty=1.3)
corrected = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(corrected)

Input prefix: (none — empty string)

Acknowledgments

All credit for the original model goes to Alexander Visheratin (@visheratin). This repository only provides an ONNX INT8 quantized conversion to make the model easier to deploy in production environments (mobile, edge, browser).

Quantization and benchmarking performed as part of the Smart Desktop Keyboard Grammar Engine project.

Citation

If you use this model, please cite the original authors:

@misc{visheratin_tiny_onnx_int8,
  title = {visheratin/t5-efficient-tiny-grammar-correction — ONNX INT8 Quantized},
  note = {Quantized version of visheratin/t5-efficient-tiny-grammar-correction},
  url = {https://huggingface.co/visheratin/t5-efficient-tiny-grammar-correction},
}