HTR-Tiny 0.3.0

Outputs are synthetic reconstructions. They are a model's estimate of pre-2018-style scientific prose given an LLM rewrite. They are not the original text and not evidence of human authorship or of AI use. Do not use this model for AI-text detection, authorship attribution, grading, or academic-integrity decisions, and do not present its output as human-written.

Part of UQO Working Paper WP-AI-0002, "Human Text Reconstruction: Learning the Inverse of LLM-Induced Transformations in Scientific Writing" (Simon-Pierre Boucher, Université du Québec en Outaouais, 2026). Try it at https://htr.spboucher.ai (model id htr-tiny-0.3.0).

Field	Value
Family	htr (this paper's encoder–decoder Transformer, trained from scratch)
Parameters	33,538,945
Base model	none (trained from scratch)
Speed (Apple M4 Pro GPU, greedy)	≈0.3 s per paragraph (greedy or beam 3, Apple GPU)
Exported	2026-11-10T06:45:15Z

What it does

Given a paragraph of scientific prose that a large language model has rewritten, the model produces a passage closer to the register of the pre-2018 human original: it reverses part of the LLM's lexical and stylistic transformation while keeping numbers, citations (as [[C1]]/⟦C1⟧ placeholders), equations and URLs unchanged. Gains are measured against the identity baseline (returning the input unchanged), which is itself a strong baseline because reconstruction from a paraphrase is non-identifiable.

Results (paired differences vs identity, chrF++, ± 1.96 SE)

Split	Δ chrF++ vs identity (B0)
Validation (3 000 pairs)	—
Test (15 649 pairs) — confirmatory	+0.44
Held-out disciplines (physics, earth sciences; 2 493)	-0.10
Held-out period (books < 1900; 2 574)	-0.53

Preservation on the test split: number recall 0.991, placeholder (citation/equation/URL) exactness 0.992. Pangram 4.1 detector sensitivity (54 test passages; input rewrites 46 % flagged): 54/54 labelled Human, mean AI-likelihood 0.0. This measures one commercial detector's behaviour on a 54-passage sample. A low AI-likelihood is not evidence of human authorship, and HTR must not be used to evade AI-text detection or for integrity decisions.

Training data

Medium stage of WP-AI-0002: denoising pretraining on 258 264 rights-verified pre-2018 passages (PMC Open Access, CC BY economics and finance journals, arXiv, public-domain books), then fine-tuning on 72 698 validated Haiku rewrite → original pairs of 14 000 passages (32k SentencePiece). The originals are rights-verified scientific passages published before 2018 (PMC Open Access, CC BY economics and finance journals, arXiv, public-domain books, NCBI Bookshelf, OpenStax); the rewrites were produced by Claude Haiku 5.5 (Anthropic) with five prompt variants. The pairs are not released.

How to use

pip install torch sentencepiece safetensors rapidfuzz sacrebleu rouge-score numpy pydantic
huggingface-cli download spboucher04/htr-tiny --local-dir htr-tiny

import sys; sys.path.insert(0, "htr-tiny")          # the repo ships the `htr` inference package
from htr.inference.pipeline import HTRReconstructor
rec = HTRReconstructor.from_pretrained("htr-tiny")
r = rec.reconstruct("LLM-rewritten paragraph …", metadata={"discipline": "economics", "pub_year": 2012, "section_role": "results"})
print(r.reconstruction); print(r.warnings)       # warnings list any number / citation / negation change

Limitations and intended use

• Research on how LLM rewriting transforms scientific prose, stylometric analysis, benchmarking inverse-paraphrase models, and drafting suggestions that must be checked by the author. Not for AI-text detection, authorship attribution or integrity decisions.
• The model can change meaning; the inference code returns warnings when numbers, citations, negations or hedges differ from the input.
• Trained on English scientific prose of 60–400 words; other genres and languages are out of scope.
• The paper's confirmatory model M* (selected by the frozen validation rule, seed 3).
• Architecture m1 does not condition on the context fields (discipline, era, section); they are recorded with the run only.

Licence

Weights: CC BY-NC 4.0 — non-commercial research use (the training pairs include CC BY-NC sources and LLM outputs). Code (the htr package) is Apache-2.0.

Citation

@techreport{boucher2026htr,
  title  = {Human Text Reconstruction: Learning the Inverse of LLM-Induced Transformations in Scientific Writing},
  author = {Boucher, Simon-Pierre},
  institution = {Université du Québec en Outaouais},
  series = {UQO Working Papers}, number = {AI-0002}, year = {2026}
}