--- license: apache-2.0 language: - en tags: - pretraining - data-curation - evaluation - llm - common-crawl pretty_name: ProseOnlyRepair — Evaluation Analysis Scripts --- # ProseOnlyRepair — Evaluation Analysis Scripts Anonymized evaluation/analysis scripts that reproduce the tables and figures in the *Repair-First* paper (under double-blind review), the per-task results behind them (`data/`), and the implementation of the ten-stage repair pipeline (`repair_pipeline_package/`). ## Contents | File | Purpose | |------|---------| | `build_paper_tables.py` | Headline tables (per-task accuracy, Paloma BPB, repair deltas) | | `build_8variant_analysis.py` | 8-variant repair-effect-by-tier analysis (POST-only excluded) | | `build_12variant_analysis.py` | 12-variant superset including POST-only (used in appendix) | | `make_paper_figures.py` | Regenerates Figures 1 and 2 from eval JSONs | | `requirements.txt` | Python dependencies (lm-evaluation-harness 0.4.12, datasets <4.0, etc.) | | `data/results.py` | Per-task results: 27 downstream tasks and 11 Paloma corpora for each tier (HQ, MQ, LQ) and configuration (V1: surface repair; V2: surface + linguistic repair) against Raw (mean over three pre-training seeds), with noise bands, deltas and z-scores; corpus statistics, recovery yield and repair-operation shares | | `data/make_tables.py` | Writes the appendix LaTeX tables (`appendix_tables.tex`, `table_anatomy.tex`, `table_summary.tex`, `appendix_analysis_tables.tex`) from `data/results.py` | | `data/make_figures.py` | Writes the result figures (`figures/fig1_teaser*.pdf`, `fig3_heatmaps*.pdf`, `fig4_paloma_split.pdf`, `fig5_ops.pdf`) from `data/results.py` | | `repair_pipeline_package/` | Implementation of the ten-stage repair pipeline (see below) | ## Variant naming Each evaluation result directory follows `{cluster}_{tier}_new[_repaired[_upsampled]]_results`: | Token | Meaning | |---|---| | `clusterA_` | Pretrained on Cluster A (anonymous identifier) | | `clusterB_` | Pretrained on Cluster B (anonymous identifier) | | `_lq` / `_mq` / `_hq` | Low / Medium / High quality CommonCrawl tier | | `_new` | PRE: pre-repair baseline | | `_new_repaired` | POST: post-repair, token budget held by truncation | | `_new_repaired_upsampled` | POST+UP: post-repair + upsampled to original token budget | ## How to use ```bash pip install -r requirements.txt # Place evaluation result JSONs under ./eval_results/{prose,paloma}/ python3 build_paper_tables.py python3 build_8variant_analysis.py python3 build_12variant_analysis.py python3 make_paper_figures.py ``` The scripts in `data/` need no evaluation JSONs (the per-task results are in `data/results.py`); run them from the repository root, and they write to it: ```bash python3 data/results.py # win/loss tallies per tier and configuration python3 data/make_tables.py # ./appendix_tables.tex, ./table_anatomy.tex, ./table_summary.tex, ./appendix_analysis_tables.tex python3 data/make_figures.py # ./figures/*.pdf ``` The accompanying evaluation result JSONs (12 variants × prose + paloma suites) are released in the paper's supplementary materials zip. ## Repair pipeline (`repair_pipeline_package/`) The ten-stage repair pipeline in its masking-first form: mixed-content masking runs before the ftfy repair. | Path | Role | |---|---| | `pipeline.py` | `RepairPipeline`: runs the stages in order and keeps an audit trail of every change | | `stages/stage01_encoding.py` … `stage10_restore.py` | encoding detection, masking, ftfy repair, artifact removal, line-level repair, character normalization, whitespace, OCR repair, quality validation, restoring the masked segments | | `masking/` | masking of non-prose segments (DOM path, plain-text path, sentinels) | | `typed_repair/` | type-specific repair of the masked code, math and structured segments | | `quality/ensemble.py` | the six-tool quality ensemble | | `config.py` | model paths (under `$REPAIR_MODELS_ROOT`, default `/data/models`) and thresholds | | `document.py`, `run.py` | the document and audit data model; the command-line entry point | From the repository root: ```bash python -m repair_pipeline_package.run input.html -o out.jsonl # one file python -m repair_pipeline_package.run ./corpus_dir -o out.jsonl # every file under a directory python -m repair_pipeline_package.run scan.txt --source-hint ocr --http-charset utf-8 ``` ```python from repair_pipeline_package import RepairPipeline doc = RepairPipeline().run(raw_bytes) # doc.text, doc.quality_passed, doc.audit, ... ``` Output is JSONL, one document per line: `path`, `encoding`, `quality_passed`, `quality_score`, `discarded`, `discard_reason`, `n_segments`, `audit`, `text`. Every model and library a stage uses (fastText, LanguageTool, NLTK, SymSpell, ByT5, KenLM, Magika, the masking classifiers, BeautifulSoup, ftfy) is optional: when one is missing, the stage logs a warning and falls back to a heuristic or skips that step. Model files go under `$REPAIR_MODELS_ROOT` at the paths in `config.py`. The docstrings call the package `repair_pipeline`; use the folder name as above, or rename the folder. ## Notes - The analysis scripts are CPU-only Python and require no GPU. - Scripts are deterministic: identical inputs → identical outputs. - The 28-task prose suite + 11-corpus Paloma BPB panel are evaluated with `lm-evaluation-harness==0.4.12` (vLLM backend). - Decoding: greedy, max-len 256 except CoQA/SQuADv2 (512). This release is anonymous for double-blind review and will be re-released with author and affiliation metadata after the review period.