---
license: apache-2.0
language:
  - en
library_name: litert
pipeline_tag: text-generation
tags:
  - coedit
  - flan-t5
  - litert
  - tflite
  - int8
  - grammar-correction
  - text-rewriting
base_model: jbochi/coedit-base
---

# CoEdIT Base LiteRT INT8

Split encoder/decoder LiteRT (`.tflite`) conversion of
[`jbochi/coedit-base`](https://huggingface.co/jbochi/coedit-base), intended for
fully local text editing in browsers, extensions, and native LiteRT clients.
This is an independently produced conversion, not an official release from
the upstream model author or Grammarly.

## Model details

| Property | Value |
| --- | --- |
| Source checkpoint | `jbochi/coedit-base` |
| Architecture | FLAN-T5 Base fine-tuned on CoEdIT |
| Format | Split LiteRT/TFLite encoder and decoder |
| Quantization | INT8 dynamic range |
| Maximum sequence length | 256 tokens |
| Maximum configured output | 64 tokens |
| Encoder size | 109,978,728 bytes |
| Decoder size | 166,428,952 bytes |
| Graph total | 276,407,680 bytes |

The encoder state is computed once and reused during autoregressive decoder
steps. Tensor names, task prefixes, sizes, and SHA-256 checksums are recorded
in `manifest.json`.

## Files

- `encoder.tflite`: quantized encoder graph
- `decoder.tflite`: quantized autoregressive decoder graph
- `tokenizer.json`: tokenizer vocabulary and model
- `tokenizer_config.json` and `special_tokens_map.json`: tokenizer metadata
- `manifest.json`: runtime contract and integrity metadata

## Instruction prefixes

```text
Fix grammatical errors in this sentence: <text>
Make the sentence formal: <text>
Change the style to casual: <text>
Make the sentence simpler: <text>
Paraphrase the sentence: <text>
```

## Runtime integration

Load the two graphs with LiteRT, tokenize to the fixed 256-token input shape,
run the encoder once, and feed its hidden states plus the decoder token buffer
to the decoder for each generated token. Consult `manifest.json` for the exact
tensor contract and EOS/padding IDs.

## Limitations

- Browser CPU/WASM autoregressive decoding can be slow, particularly for long
  rewrites. WebGPU or a native LiteRT runtime is recommended when available.
- Dynamic-range quantization can change outputs relative to the source model.
- Output should be reviewed before use; the model may omit, add, or alter
  meaning.
- The graphs use fixed-size tensors and are not a Transformers checkpoint.

## License and attribution

The source model is published under Apache License 2.0. This converted form is
distributed under the same license; see `LICENSE` and `NOTICE`.

- Source: https://huggingface.co/jbochi/coedit-base
- Upstream base: https://huggingface.co/google/flan-t5-base
- CoEdIT paper: https://arxiv.org/abs/2305.09857