bart-base-wikilarge-simplification

facebook/bart-base fine-tuned on WikiLarge for sentence-level English text simplification. It rewrites one sentence at a time into simpler English while preserving meaning.

Trained as part of a bachelor thesis on automated simplification of everyday English web text, and served as the local model by the project's backend: https://github.com/yyvs/simple-website

Intended use

Simplifying a single English sentence of everyday prose. The companion project uses it behind a browser extension that sentence-splits page text (pysbd) and sends each sentence independently.

Out of scope: multi-sentence input, whole documents, languages other than English, and any setting where a factual error would be harmful. This is a research artifact from a student project — it hallucinates (see Limitations) and has had no human evaluation.

For whole-document rewriting, use the companion model yunvs/bart-base-dwikipedia-simplification instead.

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tok = AutoTokenizer.from_pretrained("yunvs/bart-base-wikilarge-simplification")
model = AutoModelForSeq2SeqLM.from_pretrained("yunvs/bart-base-wikilarge-simplification")

text = "Despite the multiplicity of antecedent conditions, the framework mandates uniform compliance."
inputs = tok(text, return_tensors="pt", truncation=True, max_length=64)
out = model.generate(**inputs, max_length=64, num_beams=4,
                     no_repeat_ngram_size=3, repetition_penalty=1.2)
print(tok.decode(out[0], skip_special_tokens=True))

Use max_length=64. That is the length the model was fine-tuned at, and generating at BART's 512/1024 default degrades output. Input past 64 tokens is truncated, not chunked — split into sentences first.

Training

Base model	facebook/bart-base
Dataset	WikiLarge, ~117k training pairs after filtering corrupted rows
Epochs	5 (best checkpoint selected at epoch 4)
Batch size	16
Learning rate	3e-5
Weight decay	0.01
Max length	64 tokens (WikiLarge p90 ≈ 52)
Seed	42
Selection	load_best_model_at_end, metric_for_best_model="loss", early stopping (patience 4 quarter-epoch evals)
Final validation loss	0.4568

Early stopping did not halt the run; epoch 5 (val loss 0.4590) did not improve on epoch 4, and load_best_model_at_end selected epoch 4 correctly.

Evaluation

Evaluated on the ASSET test split (359 sentences, 10 human references each) with beam search (num_beams=4), against the un-fine-tuned facebook/bart-base as a zero-shot baseline.

	SARI ↑	BLEU	FKGL ↓
This model	37.80	67.26	8.18
facebook/bart-base zero-shot	21.34	89.89	10.02

Paired bootstrap significance test on the SARI improvement, 1000 resamples: p < 0.001.

Reading the BLEU number

BLEU going the "wrong way" here is expected, not a contradiction. BLEU is biased toward conservative, low-edit output, and ASSET's 10 references per sentence amplify that: an output that stays close to the source (as the zero-shot baseline largely does) has many more chances to match some reference's n-grams, inflating BLEU without the output being a good simplification. SARI — which rewards appropriate edits against the source — and FKGL both favour the fine-tuned model by a wide margin (+16.5 SARI, −1.85 grade levels). Treat BLEU here as a documented counterexample of its known limitation for simplification, not as evidence the baseline is better.

Limitations

• Hallucination. Short or fragmentary inputs can produce unrelated text; the phrase "Other websites" (a Simple English Wikipedia section heading present in the training data) has been observed as output for several unrelated short inputs. The companion backend runs a quality guard that falls back to the original text in this case.
• Conservative rewriting. Observed output often drops a subordinate clause rather than rephrasing dense vocabulary — it does not reliably simplify words like "multiplicity" or "heterogeneous entities."
• No human evaluation. All reported numbers are automatic metrics. No readability, meaning-preservation, or accessibility study has been conducted.
• Single sentences only. Multi-sentence input is out of distribution and gets truncated at 64 tokens.
• English only, and trained on Wikipedia prose — encyclopedic register, with whatever demographic and topical bias that corpus carries.

Training data provenance and licence

WikiLarge is derived from English Wikipedia and Simple English Wikipedia, which are licensed CC BY-SA. This model is released under CC BY-SA 4.0 accordingly.

WikiSmall was evaluated as an alternative and deliberately rejected: it carries baked-in named-entity anonymization artifacts unsuitable for training.

Citation

Produced for a bachelor thesis (2026). Please cite the project repository: https://github.com/yyvs/simple-website