manvadariya1/deshdraft-context-spelling-small-001

DeshDraft contextual spelling research model

User-fine-tuned byte-level sequence-to-sequence spelling corrector. This is not a next-word/next-phrase completion model. The task prefix is correct spelling: ; supply the surrounding sentence, at most 320 UTF-8 bytes. It was trained to preserve wording while correcting character/spacing errors. Preservation and user intent are not guaranteed.

Model and data

Base: google/byt5-small, revision 68377bdc18a2ffec8a0533fef03b1c513a4dd49d, Apache-2.0. The repository contains merged safetensors weights and byte-tokenizer configuration. No uploaded Python source, pickle training state, credentials, private app settings or raw private typing are required to load it. Use Transformers with trust_remote_code=False and use_safetensors=True.

Training used the frozen public Tatoeba English CC0 snapshot, with source-family splits before synthetic corruption, plus clean-copy examples. See source-manifest.json, experiment.json and the included licenses. The accompanying optional keyboard/ngram research layer used WordNet vocabulary; retain its license when distributing dictionary-derived artifacts.

Reported held-out results

These are the supplied training export's 200 public/synthetic test cases: 100 clean and 100 noisy. They are not independent population accuracy estimates or evidence of superiority to Windows, Gboard or other products.

Policy	Noisy sentences corrected exactly	Clean sentences preserved
Guarded neural	49 / 100	100 / 100
Keyboard + guarded neural	59 / 100	99 / 100

The export reports guarded neural edit precision about 0.868 and F0.5 about 0.760. Unthresholded suggestions can be wrong, including on already-correct text. Do not automatically apply all generated output. The reported T4 median neural evaluation latency is about 960 ms and p95 about 3,544 ms; these are not laptop CPU or Windows timings.

Intended desktop integration

An optional advanced contextual spelling reviewer with explicit download, bounded context, unchanged-text competition, stale-field/cancellation checks and local-only inference. Existing phrase completion remains separate. ONNX/INT8 deployment artifacts, when added, need their own validation; original neural score thresholds cannot certify a changed candidate set or quantized model. This model is experimental, not a publishable autocorrect quality guarantee.

No paid hosted inference is required or enabled by this repository. App-facing labels may show only download size and compatible hardware; this card retains technical attribution for reproducibility.

Windows candidate reviewer

runtime/context-ranker-fp32.onnx shares the context encoder across up to four candidate targets, including unchanged text. It is a CPU candidate scorer, not an autoregressive completion generator. FP32 matched the original teacher-forced objective within 0.000001 in four initial cases. INT8 changed a candidate preference in those checks and is not distributed. The Windows option remains experimental, uses a conservative margin and is off by default. See runtime parity and quality reports; these are bounded fixtures, not a product accuracy guarantee.