Repository Public Research preview
1---2language: en3license: apache-2.04library_name: htr5tags: [text2text-generation, scientific-writing, style-reconstruction, research]6pipeline_tag: text2text-generation7---89# HTR-Tiny-pilot-rc0 (v0.1.0)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 | `m1` (encoder–decoder Transformer; m3 = dual-latent content/style with copy gate) |18| Parameters | 27.4 M |19| d_model / layers (enc+dec) | 384 / 4+4 |20| Vocabulary (SentencePiece unigram) | 16000 |21| Max sequence length | 640 tokens |22| Style classes (m3) | n/a |23| Training run | `pilot_m1_tiny_denoiseinit` |24| Exported | 2026-11-09T19:25:42Z |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/HTR-Tiny-pilot-rc0")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/HTR-Tiny-pilot-rc0 --text "…" --discipline economics --year 2012`.4041## Training data42Pilot: 5 383 validated Claude Opus 5.5 rewrite pairs of 974 PMC Open Access passages (CC0/CC BY/CC BY-NC/CC BY-NC-SA/public domain), after denoising pretraining on 17 k train-split PMC passages; research candidate, NOT the release model4344## Evaluation (document-safe test split; identity baseline reported alongside)45```json46{47 "chrf": {48 "mean": 63.6045,49 "ci": [50 62.1976,51 64.86152 ]53 },54 "edit_sim": {55 "mean": 0.5777,56 "ci": [57 0.5582,58 0.595559 ]60 },61 "rougeL": {62 "mean": 0.5996,63 "ci": [64 0.5795,65 0.616966 ]67 },68 "exact_match": {69 "mean": 0.0,70 "ci": [71 0.0,72 0.073 ]74 },75 "number_recall": {76 "mean": 0.9308,77 "ci": [78 0.912,79 0.950380 ]81 },82 "placeholders_exact": {83 "mean": 0.8845,84 "ci": [85 0.8303,86 0.938187 ]88 },89 "_baseline_B0_identity": {90 "chrf": 71.2394,91 "edit_sim": 0.6665,92 "rougeL": 0.6898,93 "exact_match": 0.0,94 "number_recall": 1.0,95 "placeholders_exact": 1.096 }97}98```99Metrics: chrF++, normalised edit similarity, ROUGE-L, exact match, number recall, placeholder (citation/math/URL)100preservation, stylometric Wasserstein distance to originals. Confidence intervals are document-clustered bootstrap.101102## Limitations and risks103- Reconstruction from a paraphrase is non-identifiable: many originals map to the same rewrite; the model returns one104 plausible estimate. Lexical overlap with the true original is bounded.105- Factual integrity is enforced heuristically (protected spans, number checks); residual errors are flagged in106 `warnings`, not guaranteed absent. Always verify numbers, citations and negations.107- Trained on Claude rewrites of CC-licensed biomedical, economics/finance, arXiv and public-domain book passages;108 behaviour on other rewriters, languages or genres is reported in the paper's generalisation experiments only.109- Memorisation: training originals were deduplicated and the model is evaluated for nearest-neighbour overlap; see110 the paper. Do not use to recover specific copyrighted texts.111112## Licence113Code and weights: Apache-2.0. Data licences: see `LICENSE-MODEL.txt` and the data card. Attribution to NLM/PMC,114arXiv, publishers and Project Gutenberg sources is in the data card.115116## Citation117Boucher, S.-P. (2026). *Human Text Reconstruction: Learning the Inverse of LLM-Induced Transformations in Scientific118Writing.* UQO Working Paper AI-0002.119