spb/wpai0002_uqo

1% ============================================================================2% Appendices3% ============================================================================4\providecommand{\evid}[1]{\texttt{\detokenize{#1}}}5\newenvironment{promptblock}{\begin{quote}\small\ttfamily\raggedright}{\end{quote}}67\appendix89\section{Rewrite Prompts (version \texttt{rw-1.0})}10\label{app:prompts}1112The prompts below are reproduced verbatim from \evid{src/htr/rewrite/prompts.py}. Two typographic substitutions13are made for this document only. (i) In the code, placeholders are delimited by the mathematical white square14brackets U+27E6 and U+27E7 (for example, the citation placeholder is the letter C and the index 1 between these two15glyphs); here they are written \texttt{[[C1]]}, \texttt{[[M1]]} and \texttt{[[U1]]}. (ii) LaTeX special characters16(\texttt{\%}, \texttt{\$}) are escaped. Line breaks inside a string are not significant.1718\paragraph{System message (all variants).}19\begin{promptblock}20You rewrite passages of scientific writing for a research study on how language models transform prose. You never21invent information. Placeholders of the form [[C1]], [[M1]] and [[U1]] stand for citations, equations and links;22copy each of them verbatim, exactly once per occurrence in the source, in the position where it belongs.23\end{promptblock}2425\paragraph{User message template.} The user message is the variant instruction, a blank line, the preservation26block, a blank line, and the passage in placeholder form:27\begin{promptblock}28\{instruction\}\\[0.5ex]29\{PRESERVE\}\\[0.5ex]30<passage>\\31\{passage\_text\}\\32</passage>33\end{promptblock}3435\paragraph{Preservation block (\texttt{PRESERVE}, all variants).}36\begin{promptblock}37Preserve every scientific proposition exactly: all numbers, units, statistical values, named entities, species,38genes, drugs, places, dates, citation markers (e.g. [12], (Smith et al., 2015), [[C1]]), equations and math tokens39(e.g. \$x\$, [[M1]]), URLs/DOIs ([[U1]]), negations, hedges and expressions of uncertainty. Do not add, remove or40reorder facts. Do not add commentary, headings, bullet points or notes. Keep the passage as a single paragraph in41English. Output only the rewritten passage.42\end{promptblock}4344\paragraph{Variant A (neutral paraphrase).}45\begin{promptblock}46Paraphrase the following passage in neutral academic English. Change the wording and sentence structure freely47while keeping the meaning identical.48\end{promptblock}4950\paragraph{Variant B (polished, standardised).}51\begin{promptblock}52Rewrite the following passage as polished, standardized contemporary academic prose: improve clarity, flow and53concision, use conventional modern scientific phrasing, and correct awkward constructions.54\end{promptblock}5556\paragraph{Variant C (concise).}57\begin{promptblock}58Rewrite the following passage more concisely (aim for about 70\% of the original length) without dropping any fact,59number, citation or qualification.60\end{promptblock}6162\paragraph{Variant D (structure-preserving).}63\begin{promptblock}64Rewrite the following passage sentence by sentence: keep the same number of sentences and the same order of ideas,65but rephrase each sentence in your own words.66\end{promptblock}6768\paragraph{Variant E (diverse sampling; two independent samples per passage).}69\begin{promptblock}70Rewrite the following passage in your own words as a fluent academic paragraph. Vary the wording naturally.71\end{promptblock}7273\paragraph{Held-out prompt \texttt{H1} (plain-language expert; never used in training).}74\begin{promptblock}75Rewrite the following passage for an expert reader in plain, direct language with short sentences.76\end{promptblock}7778\paragraph{Held-out prompt \texttt{H2} (editorial polish; never used in training).}79\begin{promptblock}80You are a copy-editor at a scientific journal. Edit the following passage for publication: fix style, tighten the81prose, keep the author's meaning.82\end{promptblock}8384\paragraph{Request settings.} Model \texttt{claude-opus-5-5} (held-out models \texttt{claude-sonnet-5-5},85\texttt{claude-haiku-5-5}); effort \texttt{low}; \texttt{max\_tokens} 1\,500; no temperature set; thinking at the86provider default; Message Batches; samples per variant A--D: 1, E: 2; pilot budget ceiling 30~USD87(\evid{configs/rewriting/pilot.json}). A fingerprint of the prompt set (SHA-256 over version, system message,88preservation block and variants A--E) is logged with every job.8990\section{Hyper-parameters}91\label{app:hyper}9293Table~\ref{tab:hyper} lists the training and model hyper-parameters. The ``code default'' column gives the values94in the repository at this version (\evid{src/htr/training/trainer.py}, \evid{scripts/train.py},95\evid{src/htr/models/config.py}, \evid{src/htr/models/losses.py}); where the training script overrides a library96default, both are shown. Values used for the reported models are read from the run manifests and are pending.9798\begin{table}[ht]99\centering100\caption{Hyper-parameters. Script default / trainer default where they differ.}101\label{tab:hyper}102\small103\begin{tabularx}{\textwidth}{@{}>{\raggedright\arraybackslash}X>{\raggedright\arraybackslash}p{0.3\textwidth}l@{}}104\toprule105Hyper-parameter & Code default & Reported models \\106\midrule107\multicolumn{3}{@{}l}{\emph{Optimisation}} \\108Optimiser & AdamW, $\beta = (0.9, 0.98)$, $\epsilon = 10^{-8}$ & \pending{manifest} \\109Peak learning rate & $3 \times 10^{-4}$ & \pending{manifest} \\110Weight decay & 0.01 (not on biases, norms) & \pending{manifest} \\111Warm-up steps & 300 / 500 & \pending{manifest} \\112Maximum steps & 5\,000 / 20\,000 & \pending{manifest} \\113Schedule & linear warm-up, cosine to 10\% of peak & \pending{manifest} \\114Gradient clipping (norm) & 1.0 & \pending{manifest} \\115Gradient accumulation & 1 & \pending{manifest} \\116Batch budget & 8\,192 tokens, at most 48 pairs & \pending{manifest} \\117Evaluation interval & 250 / 500 steps & \pending{manifest} \\118Early-stopping patience & 6 evaluations (validation token CE) & \pending{manifest} \\119Precision & bf16 on CUDA, fp32 on MPS & \pending{manifest} \\120Seeds & 1, 2, 3 & \pending{manifest} \\121\midrule122\multicolumn{3}{@{}l}{\emph{Model}} \\123Size design point & tiny (Table~\ref{tab:sizes}) & \pending{manifest} \\124Dropout & 0.1 & \pending{manifest} \\125Positional encoding & RoPE (self-attention) & \pending{manifest} \\126Maximum source / target length & 512 / 512 tokens & \pending{manifest} \\127Tokenizer & SentencePiece unigram, train split only; design vocabulary 32\,000 & \pending{vocabulary size} \\128Copy gate & on & \pending{manifest} \\129Style tokens $T_s$ (M3) & 4 & \pending{manifest} \\130Style classes $K$ (M3) & (discipline, era) buckets with $\geq 20$ training pairs, plus ``other'' & \pending{$K$} \\131Gradient-reversal $\lambda$ (M3) & 0.1 & \pending{manifest} \\132\midrule133\multicolumn{3}{@{}l}{\emph{Loss}} \\134Label smoothing $\varepsilon$ & 0.1 (0 in overfit test and validation) & \pending{manifest} \\135Contrastive temperature $\tau$ & 0.07 & \pending{manifest} \\136Loss weights & Table~\ref{tab:loss-weights} & \pending{manifest} \\137\midrule138\multicolumn{3}{@{}l}{\emph{Decoding and evaluation}} \\139Decoding & greedy (beam width chosen on validation) & \pending{freeze record} \\140Bootstrap replicates & 1\,000 (development) / 10\,000 (confirmatory) & \pending{freeze record} \\141\bottomrule142\end{tabularx}143\end{table}144145\section{Reproducibility Checklist}146\label{app:checklist}147148The checklist follows the structure of the ACL Responsible NLP Research checklist. ``Yes'' means the item exists in149the repository at this version; ``Partial'' that it exists for the completed gates only; ``Pending'' that it150depends on runs not yet made.151152\begin{center}153\small154\begin{tabularx}{\textwidth}{@{}>{\raggedright\arraybackslash}X>{\raggedright\arraybackslash}p{0.1\textwidth}>{\raggedright\arraybackslash}p{0.33\textwidth}@{}}155\toprule156Item & Status & Where \\157\midrule158\multicolumn{3}{@{}l}{\emph{A. Claims, limitations and risks}} \\159Main claims stated and scoped (no first-inversion, authorship or detection claim) & Yes & Sections~\ref{sec:intro}, \ref{sec:related} \\160Limitations discussed & Pending & Limitations section (after results) \\161Potential risks and misuse discussed & Yes & Section~\ref{sec:ethics} \\162\midrule163\multicolumn{3}{@{}l}{\emph{B. Scientific artefacts}} \\164Sources of data cited and described & Yes & Section~\ref{sec:data} \\165Licence of every document recorded and respected & Yes & licence policy, per-document metadata \\166Intended use of artefacts stated & Partial & Section~\ref{sec:ethics}; model card template \\167Personal or offensive content checked & Partial & published scholarly text only; no systematic screen \\168Documentation of artefacts (data card, model card) & Pending & model-card template exists \\169Dataset statistics (counts, splits) reported & Partial & pilot corpus and splits; Medium stage \pending{counts} \\170\midrule171\multicolumn{3}{@{}l}{\emph{C. Computational experiments}} \\172Parameter counts, compute budget, infrastructure & Partial & design points (Table~\ref{tab:sizes}); per-run counts \pending{manifests} \\173Experimental setup and hyper-parameters & Partial & Section~\ref{sec:design}, Table~\ref{tab:hyper} \\174Descriptive statistics with error bars over seeds & Pending & 3 seeds, clustered bootstrap planned \\175Packages and versions for pre-processing and evaluation & Yes & \evid{pyproject.toml}; versions logged per run \\176Code for data acquisition, training and evaluation & Yes & \evid{src/htr/}, \evid{scripts/} \\177Random seeds, data hashes, run manifests & Yes & per-run manifests; split seed 20261109 \\178Pre-registration of confirmatory tests & Partial & draft v0.1; freeze \pending{commit} \\179\midrule180\multicolumn{3}{@{}l}{\emph{D. Human annotators and participants}} \\181Instructions, recruitment, payment, consent & Pending & protocol in Section~\ref{sec:design} \\182Ethics review & Pending & UQO \pending{approval} \\183\midrule184\multicolumn{3}{@{}l}{\emph{E. Use of AI assistants and paid APIs}} \\185Model identifiers, prompts and settings of the rewriting LLM & Yes & Section~\ref{sec:rewrite}, Appendix~\ref{app:prompts} \\186API spend disclosed & Yes & 29.68 USD (Table~\ref{tab:pilot-jobs}) \\187Use of AI assistance in research and writing disclosed & Yes & Section~\ref{sec:ethics} \\188\bottomrule189\end{tabularx}190\end{center}191