README.md

---
license: apache-2.0
base_model: pszemraj/flan-t5-large-grammar-synthesis
pipeline_tag: text2text-generation
library_name: transformers.js
tags:
  - grammar
  - grammar-correction
  - onnx
  - transformers.js
  - gwishy
---

# Gwishy Deep Check Pro — flan-t5-large-grammar-synthesis (ONNX, 4-bit)

ONNX conversion of
[pszemraj/flan-t5-large-grammar-synthesis](https://huggingface.co/pszemraj/flan-t5-large-grammar-synthesis)
(783M parameters) packaged for [transformers.js](https://huggingface.co/docs/transformers.js),
built for the **Deep check Pro** tier of the
[Gwishy writing assistant](https://github.com/PolymathMarsh/Gwishy).

- `onnx/*_q4.onnx` — **MatMul 4-bit weight-only** (block 64, symmetric).
  Activations stay fp32, so correction quality holds where naive dynamic
  int8 quantization of this model family hallucinates. ~688 MB total.
  This is what Gwishy loads (`dtype: "q4"`).
- `onnx/*_int8.onnx` — dynamic int8, included for comparison (~819 MB).

Conversion pipeline: `optimum-cli export onnx` (task
`text2text-generation-with-past`), decoder merge via `optimum.onnx.merge_decoders`,
quantization via onnxruntime `MatMulNBitsQuantizer` — see
[`scripts/convert-pro-model.py`](https://github.com/PolymathMarsh/Gwishy/blob/main/scripts/convert-pro-model.py).

## Usage (transformers.js)

```js
import { pipeline } from '@huggingface/transformers';
const fix = await pipeline('text2text-generation', 'MOTU01/gwishy-deep-pro', { dtype: 'q4' });
const out = await fix('Furthermore, the teams research proposal was deferred.', { max_new_tokens: 192 });
// -> "Furthermore, the team's research proposal was deferred."
```

## License

The base model is dual-licensed by its author
([Apache-2.0 / CC-BY-NC-SA-4.0](https://huggingface.co/pszemraj/flan-t5-large-grammar-synthesis));
this conversion is distributed under the Apache-2.0 option. All credit for
the model itself to [@pszemraj](https://huggingface.co/pszemraj).