cive202/humanize-ai-text-bart-large

Fine-tuned BART-large (facebook/bart-large) for AI → Human rewriting (“humanization”). This model is designed for constrained rewriting: preserve meaning while shifting style toward human-authored text.

• Architecture: encoder–decoder (seq2seq)
• Parameters: ~406M
• Task format: humanize: {ai_text} → {human_text}

📄 Paper

“Rewriting the Machine: Encoder-Decoder vs. Decoder-Only Transformers for AI-to-Human Text Style Transfer”
Authors: Utsav Paneru et al.
arXiv: https://arxiv.org/abs/2604.11687v1
Status: Preprint (2026)

Citation

@misc{paneru2026makesoundlikehuman,
      title={Please Make it Sound like Human: Encoder-Decoder vs. Decoder-Only Transformers for AI-to-Human Text Style Transfer}, 
      author={Utsav Paneru},
      year={2026},
      eprint={2604.11687},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.11687}, 
}

Quickstart

pip install -U "transformers>=4.40.0" torch sentencepiece

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_id = "cive202/humanize-ai-text-bart-large"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

ai_text = "Large language models often produce fluent, structured prose with recognizable regularities..."
inputs = tokenizer("humanize: " + ai_text, return_tensors="pt", truncation=True)

out = model.generate(
    **inputs,
    max_new_tokens=256,
    num_beams=4,
)

print(tokenizer.decode(out[0], skip_special_tokens=True))

Training summary (from project config)

Full fine-tuning (no adapters) with a standard seq2seq cross-entropy objective:

• LR / schedule: 5e-5, cosine scheduler
• Warmup ratio: 0.1
• Precision: bf16
• Effective batch size: 16 (per_device_train_batch_size = 2, gradient_accumulation_steps = 8)
• Epochs: 5
• Checkpoint selection: best checkpoint by validation loss

Dataset

Parallel chunk pairs created via sentence-aware chunking:

• Train: 25,140 pairs
• Validation: 1,390 examples
• Test (evaluation subset): 1,390 examples

Preprocessing details (high-level):

• sentence tokenization (NLTK)
• greedy packing to a token budget (≤200 tokens measured with BART-base tokenizer)
• drop pairs with fewer than 10 words on either side
• document-disjoint splits (no doc_id overlap between splits)

Evaluation (test n = 1,390)

All metrics computed on the same 1,390-example test subset.

Reference similarity (higher is better)

• BERTScore F1: 0.9240
• ROUGE-L: 0.5657
• chrF++: 55.9219

Fluency proxy

• GPT-2 perplexity (output): 27.1481
• GPT-2 perplexity (human reference): 23.6912

Linguistic marker shift (style movement)

• Mean directional marker shift: 0.8289

Qualitative note:

• This run is characterized by comparatively precise targeting of human marker means on several features (e.g., average word length and lexical diversity were extremely close to human reference means in the project’s analysis).

Limitations

• This model optimizes reference similarity and controlled rewriting; it may not “push style” as aggressively as decoder-only models that can overshoot.
• No guarantee of bypassing AI detectors.
• Generalization depends on domains/styles present in training data.

Research paper (unpublished)

Part of an unpublished manuscript (2026):

“Rewriting the Machine: Encoder-Decoder vs. Decoder-Only Transformers for AI-to-Human Text Style Transfer”

• Status: not published yet
• Link: (add your PDF/arXiv link when available)

License

MIT is a placeholder here—set this repo’s license to what you intend to distribute under, consistent with the base model’s terms.