Learning from Experience
A guide to the writing data
Understand what the sample contains, how the records connect and what your team could explore.
What you are exploring
This sample follows selected stages in the writing of ImitSAT, an AI research paper. It brings together drafts, working notes and revisions, with commentary on the writing decisions they illustrate.
21 manuscript excerpts · 24 writing questions · 3 histories · 7 skill views. The guided route follows 14 selected questions. Different questions can share passages; these are not independent training examples or an estimate of the amount of data a typical paper provides.
The histories explore contribution framing, literature argument and careful revision. Writing is the supplied example. Broader research, software and game-development histories are planned or potential coverage, not material included here.
Three different layers
| Layer | What it contains |
|---|---|
| Manuscript history | Source passages, working notes, recorded draft revisions and connections between selected states. |
| Writing commentary | Questions, skill associations and interpretations of why a change may matter. These are not original instructions or verified accounts of the writer's intention. |
| Possible tasks | Newly authored instructions and proposed review criteria showing how the records could be used. They are not historical prompts or completed model evaluations. |
Start with a passage, a question or a task
Follow the guided presentation for one writing issue at a time. Open a source passage for the full excerpt, then select its data to inspect the structured record. A passage may be connected to several writing questions.
How the text is prepared
Scientific wording, citation keys, comments and historical spellings are preserved. Private contributor labels use generic markers. The paper remains identifiable; the generic markers do not tell you who made a particular edit.
The selection does not include the complete manuscript history. In particular, the literature story starts with a partly organized draft rather than the earlier citation scratchpad.
Draft revision numbers come from Overleaf, the manuscript editor. They identify recorded states, not a person or the time spent writing.
What the sample does not establish
The published final paper, original full archive, cited works, referenced figures, experiment logs and original AI conversations were not independently inspected or reconstructed for this sample. Individual edit authorship is unknown.
Scientific assertions in the manuscript have not been independently verified. A later revision is a record of what happened, not necessarily a better answer. No model performance, reader improvement, publication outcome or buyer demand has been measured from these records.
Trying a writing task
Each task has its own input file. Proposed review criteria and selected later drafts are provided separately. The instructions were written after inspecting the history; these are example tasks, not a hidden-answer evaluation.
| Task | Input |
|---|---|
| Scientific positioning | Download task input |
| Literature argument | Download task input |
| Careful revision | Download task input |
Two inputs in the careful-revision task leave out comment-only lines. The full historical excerpts retain those comments. The task states this distinction explicitly.
Technical guide to the JSON fields
JSON is a structured text format. The field names below let a developer move from a writing question to the exact passages that support it.
| Field | How to use it |
|---|---|
evidence | Source records indexed by reference code. text contains the excerpt with private contributor labels replaced where needed. manuscript_revision is the recorded manuscript revision. |
chapters[].steps[] | Writing questions. reading_view holds the explanation and context; focus_passages contains exact selections from the source excerpts. |
trajectory_cases | Selected milestones and what earlier states add to the understanding of a later passage. |
skill_profiles | Overlapping associations between writing skills and questions. They do not create additional independent source data. |
writing_action_tasks | An index of the possible writing tasks. The separate files contain complete task inputs and review material. |
scope, counts, release, preprocessing, origin, limits | Information about the selected coverage, text preparation and use conditions. |
Codes such as A02 and T1-02 are stable links between records. They are not people, draft numbers or sample counts. Gaps in those codes reflect the selection and do not imply missing downloads.
Character offsets use Unicode code points, with an inclusive start and exclusive end. In Python use text[start:end]; in JavaScript use Array.from(text).slice(start,end).join(''). Checksum fields describe the delivered source text, not an unmodified original private archive.
Manuscript source records use text, manuscript_revision and source_ref.text_file. Task source records use the same text and revision names. The top-level version describes the data format, not a manuscript revision. Technical field names and reference codes connect the passages, writing questions, and tasks. The guided presentation uses descriptive labels instead.
Discussing further use
The presentation works offline and has no trackers, accounts or automatic message submission. A copy of the HTML contains the complete sample data, including later reference material.