What the LPCG framework generated from six JSS1 mathematics lessons, and how four blinded raters and a model judge rated it: the inputs as frozen, every run with its record, the documents the raters received and returned, and the ratings.

This is one of four datasets released with the LPCG framework from the MSc study Design and Evaluation of a Lesson-Plan-Driven Framework for Curriculum-Constrained Generation and Personalisation of Learning Materials: A Case Study of Nigerian Junior Secondary Mathematics (University of Lagos, 2026). The lessons themselves are in ng-jss1-math-lesson-plans.

3,820 files, 338.5 MB.

The study in brief

Part 1 — class revision material. For each of the six lessons the framework generated revision packs under four conditions, C0 to C3 — a ladder in which each condition adds to what the generator is given — with three draws per condition: 72 packs, generated by deepseek-v4-flash in run c0c3-prism-objective. Each pack went to two of the four raters, who scored it against each of the lesson's objectives: 144 documents and 1,212 scores. A model judge, claude-sonnet-4-6, scored the same packs in two passes: 576 scores.

Part 2 — student-specific revision material. For the three lessons with class work, each learner's case sets two versions side by side: P0, the Part 1 C3 pack from draw 1 unchanged, and P1, that same pack personalised for the learner by deepseek-v4-flash in run h2-deepseek-clean. There are 26 learner–lesson cases, each compared by two raters: 52 documents issued, 51 returned with comparisons, 199 comparisons.

path	what it holds
`inputs/eval_set.json`	the six lesson plans as frozen for Part 1, each with its digest
`inputs/frozen/`	a superseded version of one plan, kept so that runs made before its correction still resolve, with a README saying what changed
`runs/h1/`	all 105 Part 1 run directories; `MANIFEST.md` says which is reported and what the rest record, and `README.md` explains why three share one label
`runs/h1-wire/`	the request sent to the generator under each context version, and what changed between versions
`runs/h2/`	the Part 2 runs: `h2-deepseek-clean`, which is reported, and two development runs
`ratings/delivered/`	the 204 documents as the raters returned them, by part and rater, with the record of how they were built
`ratings/allocation/`	which packet went to which rater, and in what order
`ratings/part1_ratings.csv`	the 1,212 Part 1 scores, one row each, with the returned document it was read from
`ratings/part2_judgements.csv`	the 199 Part 2 comparisons, one row each, with the returned document it was read from
`ratings/documents.csv`	the 196 returned documents, one row each: how much was answered, the date the rater wrote, and their closing comment on the set
`ratings/workbooks/2026-09-13-ratings.xlsx`	the workbook the scores and comparisons were entered into, with the learner table, the judge's scores and the generation record

| ratings/instrument/ | the guide the raters worked from, and the pilot worksheets | | PROVENANCE.json | every file with its SHA-256, the repository commits it was read from, and which files differ from the study's record |

The rating tables

ratings/part1_ratings.csv and ratings/part2_judgements.csv hold the scores and comparisons of ratings/workbooks/2026-09-13-ratings.xlsx, the workbook as it stood when rating closed, one row each. ratings/documents.csv holds its Coverage sheet, one row per returned document, with the date the rater wrote on the set and the comment they left on its last page, which belong to the whole set rather than to any one score. Each row adds the path of the document the rater returned, so every rating can be checked against the page it was written on. The tables are built from the workbook and the documents and add nothing to them.

Blinding

Every document a rater received carried only an external packet id. Condition labels, lesson codes, learner ids, run labels and internal ids were kept off the page, and both the issued and the returned documents were scanned for them. In Part 2, which of P0 and P1 appeared first alternated.

What differs from the study's record

83 of the 3,820 files differ from the study's own record, and PROVENANCE.json marks each one.

The school is not named. Where the study's plans carried its abbreviation, this release has sch.
The digests that pin each plan — plan_sha256 in inputs/eval_set.json and in each run record, plan.<lesson> in each run's fingerprint, and the superseded plan's file name — are digests of the plans as published, not of the study's originals, which would identify the school. 14 digests were replaced. A plan published in ng-jss1-math-lesson-plans verifies against its digest here, and so does a plan in inputs/ once its sets_work judgements are set aside, since those were added after the digest was taken. Digests of plan versions published nowhere, and run fingerprint ids, are as recorded.

Consent and privacy

Raters appear only as rater-1 to rater-4, learners only as student-01 to student-09, and no learner is named in any document or comment. The Part 2 documents carry a learner's transcribed class work, because a rater had to see it to judge the personalisation; it is released under the same written parental consent as ng-jss1-math-lesson-plans.

Limitations

Six lessons from one teacher and one class; Part 2 covers three lessons and at most nine learners; four raters.
Generated material is published as generated, errors included. Some packs contain literal HTML tags where formatting was escaped.
Runs identify the code that produced them by digest; the code is not part of this dataset.
Results are reported in the dissertation, not here: this dataset holds the ratings they were computed from.

Licence and credit

CC BY 4.0. The reported runs were generated by deepseek-v4-flash, and development runs record the models they used. The judgements are the four raters' and claude-sonnet-4-6's.

Related datasets

ng-jss1-math-lesson-plans: the six lessons these packs were generated from
ng-jss1-math-curriculum: the curriculum and scheme of work the lessons follow
ng-jss1-math-textbook-linkage: an open JSS1 textbook linked to the curriculum's objectives

Citation

@mastersthesis{amuda2026lpcg,
  author = {Amuda, Tosin},
  title  = {Design and Evaluation of a Lesson-Plan-Driven Framework for Curriculum-Constrained Generation and Personalisation of Learning Materials: A Case Study of Nigerian Junior Secondary Mathematics},
  school = {University of Lagos},
  type   = {MSc dissertation},
  year   = {2026}
}