Free Paraphrasing Tool — No Signup, Sources Cited

Rewriting text well is a generative task, so this page is honest about which parts are reproducible and which are not. The wording comes from a neural model and will vary; the change metrics underneath are pure arithmetic that you can reconcile by hand. This is the same split our text summarizer uses, where the TextRank scores are the verified part.

1. Sentence split and quote masking

The input is split into sentences with Intl.Segmenter (falling back to a terminal-punctuation regex). When “Preserve quoted text” is on, anything inside matched double or single quotes is swapped for a placeholder the model copies through, then restored verbatim afterwards — so a quotation you intend to cite is never silently reworded.

2. The rewrite (server-side)

Each sentence is sent to humarin/chatgpt_paraphraser_on_T5_base, a T5-base model fine-tuned on ChatGPT-generated paraphrase pairs (Apache-2.0), through the Hugging Face Inference API from a server-only route. The browser never downloads model weights. The four modes map onto the decoding parameters defined in the Hugging Face generation docs:

• Standard: diverse beam search (5 beams, 5 groups).
• Fluent: sampling (top_p 0.95, temperature by strength).
• Formal: beam search (5 beams), prefix paraphrase formally:.
• Simple: beam search (4 beams), prefix paraphrase in simple english:.

Synonym strength raises or lowers the temperature and diversity penalty — Light (temp 0.5), Medium (temp 0.7), Strong (temp 1.0). Higher strength swaps more words but risks drifting further from the original meaning, which the metrics will flag.

3. The deterministic metric pass

Once the rewrite returns, the page computes a set of figures that are fully reproducible from the before/after pair:

• Word overlap — the Jaccard index of the two content-word sets:

               |A ∩ B|
  overlap % = ─────────── × 100
               |A ∪ B|

  where A and B are the content words (after lowercasing and removing the 126-word stop-word list) of the original and the rewrite. Lower means more reworded.
• Words changed — original content words, counted in order, that do not appear in the rewrite, over the total content words.
• Length — word and sentence counts, original versus output.
• Reading grade — Flesch–Kincaid grade of each text and the delta, using the same implementation as our readability checker.

The overlap is computed two independent ways — once with the |A|+|B|−|A∩B| identity and once by building the union set directly — and the page shows they agree. If they ever diverged, there would be a bug; reporting both is the same cross-check our tax calculator runs against the IRD formula.

4. Guardrails and limits

If the model returns the sentence unchanged (overlap above 95%), the server retries once at the strongest setting rather than handing back your own text. If the rewrite drops below 25% overlap, the page shows a “heavily reworded” warning so you re-check that the meaning held. Input is capped at 5,000 characters and 40 sentences per request. Text that is mostly non-Latin script is rejected with a clear message — the model is English-only and we will not fake a Sinhala or Tamil rewrite.

Because the wording varies, the reconciled numbers below are the deterministic metrics for a fixed before/after pair. The verifyWorkedExamples() function in the tool’s data module reproduces Examples A and B exactly.

Paraphrasing is a writing aid, not a way around citation. If the idea, finding, or argument belongs to someone else, you must cite the source even after rewording it — changing the words does not change who the thought belongs to. Submitting reworded material as your own original work, without attribution, is still plagiarism under every university policy in Sri Lanka and abroad.

This tool does not claim to evade Turnitin, AI detectors, or any integrity check, and you should not use it to try. Its honest use is to smooth your own awkward phrasing, rephrase a passage you then cite, or reduce accidental self-plagiarism against your earlier drafts — with the source still referenced. When in doubt, write from your own understanding and cite generously.

• Hugging Face — humarin/chatgpt_paraphraser_on_T5_base (model card)
• Raffel et al., 2020 — Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer (T5)
• Zhang, Baldridge & He, 2019 — PAWS: Paraphrase Adversaries from Word Scrambling
• Hugging Face Transformers — text generation parameters (num_beams, temperature, top_p, repetition_penalty)
• Hugging Face Inference API — official documentation

The model card, API contract, and references were last cross-checked on 2026-06-05. The page is reviewed whenever the Hugging Face Inference API contract changes or the upstream model repository is re-uploaded.