| --- |
| pretty_name: Introductory Python Paper Assessment Scripts |
| license: other |
| license_name: research-use-terms |
| license_link: LICENSE.md |
| language: |
| - en |
| tags: |
| - education |
| - assessment |
| - handwriting |
| - paper-based-testing |
| - global-south |
| - responsible-ai |
| size_categories: |
| - n<1K |
| extra_gated_prompt: >- |
| These are educational records with redacted registration fields, not guaranteed |
| anonymous handwriting. Use is limited to research and assessment evaluation. |
| Do not attempt re-identification, writer identification, student profiling, |
| surveillance, or high-stakes automated decisions. Read LICENSE.md and DATA_STATEMENT.md. |
| extra_gated_fields: |
| Affiliation: text |
| Intended use: text |
| I agree to the stated use restrictions: checkbox |
| --- |
| |
| # Introductory Python Paper Assessment Scripts |
|
|
| A distribution and numeric-reproduction artifact for John Oghenereuse Fiewor's |
| *Teacher-in-the-Loop AI Assessment for African Classrooms: Paper-Based Grading, |
| Offline Exams, and Code-Switched Oral Verification*, accepted to GlobalSouthAI |
| at NeurIPS 2026. The experiments used a research prototype, not the production |
| GradrAI application. |
|
|
| ## Contents and version scope |
|
|
| - `data/scripts/script_001.pdf` through `script_060.pdf`: 60 compact script PDFs, |
| totaling 217 pages. These are distribution derivatives, not the original |
| model-inference inputs. |
| - `data/question.pdf`, `data/guide.pdf`: question paper and marking guide. |
| - `human_scores/score_sheet_original.csv`, `score_sheet_grader2.csv`: the current |
| original-lecturer and additional-grader exports supplied by the author. IDs |
| 1–60 are study script indices linking scores to released PDFs. |
| - `human_scores/score_source_notes.md`: version history, unresolved mark |
| discrepancies, and reference construction. |
| - `analysis/data/`: earlier/current numeric labels, fixed numeric model outputs, |
| explicit scoring policy, original speech observations, audio subset manifest, |
| and provenance digests. No learner-name mapping or raw model transcript is included. |
| - `analysis/prompts/`: saved extraction and grading instructions and rubric. |
| Workstation path prefixes are replaced with `[local-course-materials]/`; |
| the remaining historical instruction text is preserved. A listed presentation |
| was missing and one PDF page lacked extractable text in the context record. |
| - `analysis/`: portable Python analysis, regression checks, and derived summaries. |
| - `metadata.csv`: file identity and PDF derivative information. |
|
|
| The submitted grading table used earlier human labels. The camera-ready analysis |
| keeps the same predictions but uses the later corrected exports. The mean absolute |
| errors change from **4.4334 / 5.2263** to **4.3892 / 5.1960**, without new inference. |
| Both snapshots remain reproducible. Do not merge them or treat either rater as |
| independently adjudicated ground truth. |
|
|
| ## Reproduce the numeric results |
|
|
| Python 3.11 or newer, standard library only, from a downloaded repository: |
|
|
| ```bash |
| python3 analysis/camera_ready_stats.py --json-out analysis/camera_ready_stats.json |
| python3 analysis/human_reference_stats.py --original human_scores/score_sheet_original.csv --grader2 human_scores/score_sheet_grader2.csv |
| python3 analysis/asr_stats.py |
| python3 -m unittest discover -s analysis -p 'test_reproduction.py' -v |
| ``` |
|
|
| These commands recompute references, finalization, statistics, and speech |
| aggregates from released observations. They do **not** call a model or rerun OCR, |
| ASR, or grading. Full inference replication additionally requires the original |
| image inputs, runtime orchestration, and access to historical commercial models; |
| that is not claimed by this release. |
|
|
| Bootstrap intervals use 10,000 script-level resamples and seed 20260901. Grader |
| observations for the same script are not treated as independent bootstrap units. |
|
|
| ## Scoring and metric cautions |
|
|
| The exam selects four questions with equal normalized weight, totaling 60. |
| Raw rubric maxima are 18, 17, 17, 17, and 24. Each question is clipped to its raw |
| cap, normalized to 15, and the best four attempted questions are summed. Therefore |
| **a raw sum above 60 is not itself an invalid normalized score**. Blank score |
| cells remain distinct from explicit zeros. |
|
|
| The current policy-scale human mean absolute difference is 5.28287 marks. |
| The previous 6.37 raw-sum comparison used a different scale and must not be used |
| to claim that AI outperformed human graders. Validation-message counts are not |
| independent measurements of semantic reasoning failures. |
|
|
| Speech summaries use the original August 11-clip and 30-clip result snapshots. |
| WER/CER are macro means of **per-clip rates capped at 1**, with empty transcripts |
| penalized by 1. Provider score gap excludes missing scores and compares two model |
| scores, not an ASR score to a human grade. The original Set A source is the |
| [report and linked JSON](https://gist.github.com/Fiewor/7a56858723b15d59f3786733e154704e). |
| The [27-clip audio repository](https://huggingface.co/datasets/fiewor/gradrai-viva-codeswitch-benchmark) |
| contains later reruns and 16 additional ASR-draft references. Its results cannot |
| replace the original evaluation. `submission_191_set_a_manifest.csv` pins the |
| 11 evaluated references to that repository's audio revision. |
|
|
| One Set B Hausa reference contains an English translation preamble rather than a |
| usable transcription. `reference_quality_flags.csv` records the issue. Summary |
| code retains the original 30-clip table and separately reports a 29-clip |
| sensitivity analysis excluding that clip for every provider. |
|
|
| ## Consent, privacy, and derivative limitations |
|
|
| The author confirms consent for the release of redacted scripts. Registration |
| fields are obscured; an additional header redaction was flattened into script 1. |
| A subsequent review inspected rendered headers of all 217 released script pages. |
| That targeted review is **not an exhaustive body-text PII audit**, nor a guarantee |
| against identifying handwriting or incidental information in answers. |
|
|
| The released pages were rasterized to a 1700-pixel long edge and JPEG-recompressed |
| for practical distribution. This changes pixels and can change OCR behavior. |
| It is **not a demonstrated anonymization technique**. Gates and use restrictions |
| also do not prevent an authorized recipient from copying files. Keep these |
| limitations in view when reusing educational records. See DATA_STATEMENT.md. |
| |
| ## Intended use and citation |
| |
| Use for educational-assessment research, document understanding, workflow audits, |
| and reproducibility studies under LICENSE.md. No biometric/writer identification, |
| re-identification, student profiling, surveillance, or consequential automated |
| grading is permitted. Preserve the distinction between raw and normalized scores, |
| current and historical references, and observed outputs versus new inference. |
| |
| Cite the paper title, author, this repository, and the exact Hub commit used. |
| Questions can be raised through this repository's Discussions interface; do not |
| post learner-identifying material there. |
| |
