Model card
HTR-Tiny (v0.3.0)
Human Text Reconstruction Transformer — a research model that maps an LLM-rewritten scientific passage back towards the distribution of the pre-2018 human-authored original. Outputs are synthetic reconstructions: they are not the original text, and they are not evidence of human authorship or of AI use. Part of UQO working paper WP-AI-0002.
| Field | Value |
|---|---|
| Architecture | m1 (encoder–decoder Transformer; m3 = dual-latent content/style with copy gate) |
| Parameters | 33.5 M |
| d_model / layers (enc+dec) | 384 / 4+4 |
| Vocabulary (SentencePiece unigram) | 32000 |
| Max sequence length | 640 tokens |
| Style classes (m3) | n/a |
| Training run | medium_m1_tiny_mediumdenoiseinit_s3 |
| Exported | 2026-11-10T06:45:15Z |
Intended use
Research on how LLM rewriting transforms scientific prose; stylometric analysis; benchmarking inverse-paraphrase models; a tool for authors who want a draft re-aligned with historical scientific register as a suggestion that must be checked. Not intended for AI-text detection, authorship attribution, academic-integrity decisions, or production of text presented as human-written.
How to use
from htr.inference.pipeline import HTRReconstructor
rec = HTRReconstructor.from_pretrained("path/to/HTR-Tiny")
r = rec.reconstruct("LLM-rewritten paragraph …", metadata={"discipline": "economics", "pub_year": 2012, "section_role": "results"})
print(r.reconstruction); print(r.warnings)CLI: htr reconstruct --model path/to/HTR-Tiny --text "…" --discipline economics --year 2012.
Training data
Medium stage: denoising pretraining on 258 264 rights-verified pre-2018 passages (PMC OA, CC BY economics/finance journals, arXiv, public-domain books; 20 000 steps), then fine-tuning on 72 698 validated (Haiku 5.5 rewrite → original) pairs of 14 000 passages; 32k SentencePiece; seed 3 (M* by the frozen validation rule). Validation metrics below are exploratory; confirmatory test results in the paper.
Evaluation (document-safe test split; identity baseline reported alongside)
| Metric | Mean | 95 % CI | Identity baseline |
|---|---|---|---|
chrf | 68.84 | [68.03, 69.61] | 68.42 |
edit_sim | 0.6786 | [0.6706, 0.6863] | 0.6785 |
rougeL | 0.6875 | [0.6789, 0.6960] | 0.6841 |
exact_match | 0.0000 | [0.0000, 0.0000] | 0.0000 |
number_recall | 0.9906 | [0.9877, 0.9934] | 1.0000 |
placeholders_exact | 0.9953 | [0.9923, 0.9979] | 1.0000 |
Metrics: chrF++, normalised edit similarity, ROUGE-L, exact match, number recall, placeholder (citation/math/URL) preservation, stylometric Wasserstein distance to originals. Confidence intervals are document-clustered bootstrap.
Limitations and risks
- Reconstruction from a paraphrase is non-identifiable: many originals map to the same rewrite; the model returns one plausible estimate. Lexical overlap with the true original is bounded.
- Factual integrity is enforced heuristically (protected spans, number checks); residual errors are flagged in
warnings, not guaranteed absent. Always verify numbers, citations and negations. - Trained on Claude rewrites of CC-licensed biomedical, economics/finance, arXiv and public-domain book passages; behaviour on other rewriters, languages or genres is reported in the paper's generalisation experiments only.
- Memorisation: training originals were deduplicated and the model is evaluated for nearest-neighbour overlap; see the paper. Do not use to recover specific copyrighted texts.
Licence
Code and weights: Apache-2.0. Data licences: see LICENSE-MODEL.txt and the data card. Attribution to NLM/PMC, arXiv, publishers and Project Gutenberg sources is in the data card.
Citation
Boucher, S.-P. (2026). Human Text Reconstruction: Learning the Inverse of LLM-Induced Transformations in Scientific Writing. UQO Working Paper AI-0002.