Repository Public Research preview
1---2language: en3license: apache-2.04library_name: htr5tags: [text2text-generation, scientific-writing, style-reconstruction, research]6pipeline_tag: text2text-generation7---89# {name} (v{version})1011**Human Text Reconstruction Transformer** — a research model that maps an LLM-rewritten scientific passage back12towards the distribution of the pre-2018 human-authored original. Outputs are **synthetic reconstructions**: they are13not the original text, and they are not evidence of human authorship or of AI use. Part of UQO working paper WP-AI-0002.1415| Field | Value |16|---|---|17| Architecture | `{arch}` (encoder–decoder Transformer; m3 = dual-latent content/style with copy gate) |18| Parameters | {n_params_m} M |19| d_model / layers (enc+dec) | {d_model} / {layers} |20| Vocabulary (SentencePiece unigram) | {vocab} |21| Max sequence length | {max_len} tokens |22| Style classes (m3) | {style_classes} |23| Training run | `{run}` |24| Exported | {exported_at} |2526## Intended use27Research on how LLM rewriting transforms scientific prose; stylometric analysis; benchmarking inverse-paraphrase28models; a tool for authors who want a draft re-aligned with historical scientific register **as a suggestion** that29must be checked. Not intended for AI-text detection, authorship attribution, academic-integrity decisions, or30production of text presented as human-written.3132## How to use33```python34from htr.inference.pipeline import HTRReconstructor35rec = HTRReconstructor.from_pretrained("path/to/{name}")36r = rec.reconstruct("LLM-rewritten paragraph …", metadata={"discipline": "economics", "pub_year": 2012, "section_role": "results"})37print(r.reconstruction); print(r.warnings)38```39CLI: `htr reconstruct --model path/to/{name} --text "…" --discipline economics --year 2012`.4041## Training data42{training_data}4344## Evaluation (document-safe test split; identity baseline reported alongside)45```json46{metrics_json}47```48Metrics: chrF++, normalised edit similarity, ROUGE-L, exact match, number recall, placeholder (citation/math/URL)49preservation, stylometric Wasserstein distance to originals. Confidence intervals are document-clustered bootstrap.5051## Limitations and risks52- Reconstruction from a paraphrase is non-identifiable: many originals map to the same rewrite; the model returns one53 plausible estimate. Lexical overlap with the true original is bounded.54- Factual integrity is enforced heuristically (protected spans, number checks); residual errors are flagged in55 `warnings`, not guaranteed absent. Always verify numbers, citations and negations.56- Trained on Claude rewrites of CC-licensed biomedical, economics/finance, arXiv and public-domain book passages;57 behaviour on other rewriters, languages or genres is reported in the paper's generalisation experiments only.58- Memorisation: training originals were deduplicated and the model is evaluated for nearest-neighbour overlap; see59 the paper. Do not use to recover specific copyrighted texts.6061## Licence62Code and weights: Apache-2.0. Data licences: see `LICENSE-MODEL.txt` and the data card. Attribution to NLM/PMC,63arXiv, publishers and Project Gutenberg sources is in the data card.6465## Citation66Boucher, S.-P. (2026). *Human Text Reconstruction: Learning the Inverse of LLM-Induced Transformations in Scientific67Writing.* UQO Working Paper AI-0002.68