| # Data dictionary |
|
|
| Every file and field in this dataset. Read `README.md` first for what the dataset is and the terms of use. |
|
|
| ## Codes |
|
|
| | Code | Means | Notes | |
| |---|---|---| |
| | `SES-####` | one coaching session: one classroom recording sent to Rumi and everything made from it | the key that joins every file | |
| | `TCH-####` | one teacher | the same teacher keeps the same code in every session they appear in | |
| | `OBS-##` | one observer (coach) in the Islamabad programme | the same coach keeps the same code | |
|
|
| The codes were made fresh for this dataset. They are not production ids and cannot be looked up anywhere else. |
|
|
| ## Folder layout |
|
|
| | Path | One file per | What it holds | |
| |---|---|---| |
| | `sessions.csv`, `sessions.parquet` | dataset | one row per session (fields below) | |
| | `metadata.csv` | dataset | the Hugging Face audio loader's index: `file_name` plus the `sessions.csv` fields | |
| | `audio/SES-####.mp3` | session | the full recording, mono MP3 at 48 kbps, re-encoded from the original | |
| | `audio_original_codec/SES-####.<ogg\|m4a\|mp3>` | session | the original audio stream exactly as uploaded (copied bit for bit); only the container's tags were removed | |
| | `transcripts/SES-####.txt` | session | the transcript exactly as the product stored it | |
| | `tokens/SES-####.json` | session | the speech-to-text engine's word-level tokens, exactly as stored | |
| | `diarization/SES-####.json` | session | the speaker segments, exactly as stored | |
| | `analysis/SES-####.json` | session | what Rumi produced from the transcript: scores, moments, report | |
| | `human/SES-####.json` | Islamabad session | the coach's own record of the same lesson | |
| | `lesson_plans/SES-####.txt` | Islamabad session with a library plan | the text of the lesson plan the coach linked | |
| | `rubrics/<framework>.md` | framework | each rubric in plain words: indicators, levels, how scores add up | |
| | `prompts/*.md` | model step | the production prompts for scoring, moments and reports (`prompts/INDEX.md` first) | |
| | `MANIFEST.csv`, `totals.json` | dataset | sha256 and size of every file; counts and hours | |
|
|
| ## `sessions.csv` / `sessions.parquet` / `metadata.csv` |
|
|
| | Field | Type | Meaning | |
| |---|---|---| |
| | `session_code` | text | `SES-####` | |
| | `teacher_code` | text | `TCH-####`, the teacher whose lesson was recorded | |
| | `observer_code` | text | `OBS-##` for the Islamabad coach who observed; empty when the teacher recorded the lesson themselves | |
| | `programme` | text | `islamabad_observe2` (Islamabad, Pakistan: a coach observes in person with a structured form while the lesson is recorded), `rumi_pakistan`, `rumi_tanzania`, `rumi_palestine`, `rumi_kenya` (teachers send their own recording to Rumi on WhatsApp) | |
| | `country` | text | Pakistan, Tanzania, Palestine or Kenya | |
| | `language` | text | the transcript language the product stored: `ur` Urdu, `sw` Kiswahili, `ar` Arabic, `en` English | |
| | `framework` | text | the rubric the session was scored against: `fico`, `hots`, `pectaa`, `oecd`, `cot`, `mewaka`, `palestine`, `tpad`; empty when no analysis was made (`totals.json` counts these as `none`). See `rubrics/` | |
| | `framework_version` | text | the rubric version stored with the analysis, where the product stored one | |
| | `prompt_version` | text | Islamabad only: the version of the prompt that picked Rumi's moments (`observe2-moments-v1`, `observe2-moments-v2`) | |
| | `observation_type` | text | as stored: `self_recorded` (the teacher recorded their own lesson) or `leader_observation` (someone else, a coach or school leader, recorded it). Every Islamabad session is `leader_observation`; one Tanzania session is too | |
| | `date_local`, `time_local`, `tz` | text | the date and time the recording was **uploaded** to Rumi, in the local time of the programme: PKT (UTC+5), EAT (UTC+3), or Palestine time with its UTC offset written in `tz`. The lesson ended at or before this time; the product does not store when recording started | |
| | `uploaded_at_utc` | ISO 8601 | the same moment in UTC, as stored | |
| | `duration_s` | number | the recording's length in seconds, as stored | |
| | `grade`, `subject` | text | where the product stored them; often empty | |
| | `has_human_record` | bool | a `human/` file exists (Islamabad sessions) | |
| | `has_lesson_plan` | bool | the product recorded that a lesson plan was linked to the session. The plan's text is shipped only where it came from Rumi's library (`lesson_plan_file`, 18 sessions); uploaded plans are not shipped, because they can carry school and teacher headers | |
| | `status` | text | the session's final status in the product: `completed` / `observer_review_complete` (feedback finished), `observe2_ready` / `observe2_checking` (Islamabad: moments made, coach's check not finished), `observe2_failed`, `abandoned`, `conducting_conversation` (the teacher stopped part-way). Whatever was produced before the session stopped is shipped | |
| | `diarization_confidence` | number | as stored by the product, on a 0 to 100 scale | |
| | `original_audio_codec` | text | file extension of the original audio (`ogg` = Opus, `m4a` = AAC, `mp3`) | |
| | `audio_file`, `audio_original_file`, `transcript_file`, `tokens_file`, `diarization_file`, `analysis_file`, `human_file`, `lesson_plan_file` | path | relative paths to the session's files; empty when that file does not exist | |
|
|
| `metadata.csv` has the same fields, with `file_name` (= `audio_file`) first, so that `load_dataset("audiofolder")` or |
| the Hugging Face loader can pair each MP3 with its row. |
|
|
| ## `transcripts/SES-####.txt` |
|
|
| One line per speaker turn, blank line between turns, exactly as the product stored it and as the scoring model read it: |
|
|
| ``` |
| [MM:SS] Teacher (UR): text |
| [MM:SS] Student (UR): text |
| ``` |
|
|
| `[MM:SS]` is the turn's start from the beginning of the recording. `Teacher` / `Student` are the product's speaker |
| labels. The bracketed code is the language of the turn. |
|
|
| ## `tokens/SES-####.json` |
|
|
| `{"session_code", "tokens": [...]}`. Each token is a piece of a word (a word may be split across tokens; a token |
| that begins with a space starts a new word): |
|
|
| | Field | Meaning | |
| |---|---| |
| | `text` | the token's text | |
| | `start_ms`, `end_ms` | position in the audio, milliseconds from the start | |
| | `speaker` | the engine's speaker number for this token (text, e.g. `"1"`) | |
| | `language` | the engine's language code for this token (text, e.g. `"ur"`) | |
| | `confidence` | the engine's confidence for this token, 0 to 1 | |
| | `is_audio_event`, `translation_status` | engine fields, as stored | |
|
|
| ## `diarization/SES-####.json` |
|
|
| As stored: `segments`, `speakers`, `totalSegments`, `confidence`, and on 51 sessions `stt` (the engine used). `speakers` groups the same segments by speaker (`id`, `label`, `segments[]`). Each segment: `text`, `label` (`Teacher` / `Student`, the label |
| used in the transcript), `speaker` (the engine's speaker number), `start_ms`, `end_ms`. |
|
|
| ## `analysis/SES-####.json` |
|
|
| Everything the product made from the recording, as stored. Production ids, storage links and phone fields are removed; |
| a teacher's or coach's full name in `coach_brief.brief` and `teacher_delivery.caption` is replaced with their code; |
| nothing else is changed. The shape differs by programme. |
|
|
| **Islamabad (`programme = islamabad_observe2`)** |
| |
| | Key | Meaning | |
| |---|---| |
| | `framework`, `prompt_version` | as in `sessions.csv` | |
| | `rumi_moments` | Rumi's moments: timestamped events picked from the transcript (`moments[]`: `id`, `minute`, `type`, `label`, `quote`, `moment`, `why`; plus `counts`, `verified` (what the checks kept, dropped or repaired), `fidelity` (lesson-plan fidelity, when a plan was graded)) | |
| | `rumi_levels` | the levels Rumi's moments alone add up to, per indicator (`level`, `source`, `because`), before the coach's check | |
| | `moments_ready_at` | when the moments were ready (UTC) | |
| | `coach_brief` | the product's record of the observation after the coach's check: the levels, the priority, the brief the coach was given | |
| | `scores`, `domains` | the FICO-17 totals per section and overall (see `rubrics/fico.md`) | |
| | `strengths`, `focus_area`, `growth_opportunities`, `also_work_on` | the feedback content | |
| | `report_narrative` | the report text sent to the teacher | |
| | `teacher_delivery` | how and when it was sent (`channel`, `status`, `caption`, `sent_at`) | |
| | `lp_fidelity` | the lesson-plan fidelity result, when a plan was graded | |
| | `prioritized_action` | the one action the teacher was asked to take next, where stored | |
| | `pipeline_times_utc` | `created_at` (upload), `transcription_started_at`, `transcription_completed_at`, `analysis_started_at`, `analysis_completed_at`, `report_generated_at`, `completed_at`, as stored | |
| |
| **Pakistan, Tanzania, Palestine, Kenya** |
|
|
| | Key | Meaning | |
| |---|---| |
| | `framework`, `framework_version` | as in `sessions.csv` | |
| | `analysis_data` | the full analysis as stored: `scores` (`overall_marks`, `overall_max_marks`, `overall_percentage`), the framework's indicator blocks (`areas`, `domains`, `goal*` or `practices`, depending on the framework: see its rubric file), `strengths`, `growth_opportunities`, `recommendations`, `notable_moments`, `executive_summary` (the report narrative), and further keys the product adds on some sessions (for example `topic`, `subject`, `language`, `performance_band`, `talk_time`, `questions`, `reflective_corpus`, `debrief_reflection`, `ai_calls`). Sessions with no framework have little or no analysis | |
| | `prioritized_action` | the one action the teacher was asked to take next, where stored | |
| | `pipeline_times_utc` | `created_at` (upload) and `completed_at`, as stored | |
|
|
| Report language: Pakistan reports are stored in Urdu and Tanzania reports in Kiswahili, with no English version. |
| Palestine reports are stored in English (`executive_summary_en`). Kenya reports follow the session language (English or Kiswahili). |
|
|
| ## `human/SES-####.json` (Islamabad only) |
|
|
| The coach's record of the same lesson, from the structured WhatsApp form the coach filled while observing. |
|
|
| | Key | Meaning | |
| |---|---| |
| | `coach_code` | `OBS-##` | |
| | `period_minutes` | the length of the class period the coach entered | |
| | `class_size_present` | the number of children present, as the coach entered it | |
| | `times` | when the form was created, opened, part 1 done, part 2 done, sealed; when the moments were ready; when the coach finished checking them (UTC) | |
| | `part1` | the coach's answers for the first part of the lesson (warm-up to guided practice), keys `p1_*` | |
| | `part2` | the coach's answers for the second part (independent work to wrap-up), keys `p2_*` | |
| | `other_answers` | the remaining form answers (lesson plan used, focus indicators, priority, incidents) | |
| | `coach_check_of_moments` | for each of Rumi's moments shown to the coach: `moment_id`, `type`, `heard` (did the coach hear it happen) | |
| | `evidence_review` | the coach's review screen: `<indicator>_final` levels, evidence the coach added, the priority | |
| | `final_levels` | the coach's final level for each of the 17 indicators: `1` to `4`, `NA` (no chance to see it), `IE` (could not tell) | |
| | `debrief` | the coach's spoken debrief with the teacher, if recorded: `recorded_at`, `completed_at`, `transcript` (as stored), `rumi_feedback_to_coach` (the product's feedback to the coach on the debrief) | |
|
|
| Indicator codes (`C1` to `C8`, `D1` to `D5`, `F1` to `F4`) and their level meanings are in `rubrics/fico.md`, section A. |
|
|
| ## `lesson_plans/SES-####.txt` |
| |
| The text of the lesson plan the coach linked, from Rumi's lesson-plan library. Present for 18 sessions. |
| |
| ## Times and zones |
| |
| Every `*_at` field inside the JSON files is UTC. Only `date_local` / `time_local` are local. |
|
|
