Real classroom lessons, recorded on a phone, sent to Rumi (Taleemabad's AI coach for teachers) on WhatsApp, and turned into feedback for the teacher. For each lesson you get the full recording, the transcript (the written text of everything said, as Rumi's speech-to-text heard it), what Rumi concluded about the lesson, and the report the teacher received. For 50 lessons in Islamabad you also get the record a human coach made while sitting in the same room.

Everything here is what the system produced and what teachers saw. We have not cleaned or corrected any of it (one exception, on names, is under "How it was prepared"). Transcripts, speaker labels, scores, moments and reports may contain errors, and so may the coaches' records. Working out where, and how much it matters, is the point. The case-study page explains every word used here; DICTIONARY.md explains every file and field.

What is in it

Programme	Country	Language	Rubric	Sessions	Teachers	Hours	Human coach record
Islamabad classroom observation	Pakistan	Urdu	FICO-17	50	25	24.1	yes
Rumi Pakistan	Pakistan	Urdu	HOTS (21), PECTAA (4), OECD (1), COT (1)	27	15	8.7	no
Rumi Tanzania	Tanzania	Kiswahili (22), English (1)	MEWAKA	23	15	8.6	no
Rumi Palestine	Palestine	Arabic	Palestine framework	8	5	3.6	no
Rumi Kenya	Kenya	English (3), Kiswahili (2)	TPAD (2), FICO 1.0 (1), none (2)	5	4	1.5	no
All	4 countries	4 languages	8 rubrics	113	64	46.5	50

Recordings run from 15 to 42 minutes; every one is the full recording as uploaded, never cut. Sessions were uploaded between 28 August and 28 October 2026. Many teachers appear in more than one session; a teacher keeps one code throughout.

How a session is made

Hear. The teacher (or, in Islamabad, a coach) records the lesson on a phone and sends it to Rumi on WhatsApp. A speech-to-text engine transcribes it, splits it into speaker turns and labels each turn Teacher or Student.
Think. A language model reads the transcript against the rubric for that programme and scores each indicator. In Islamabad, it instead picks timestamped moments from the transcript, and those moments propose levels.
Speak. The scores and moments become a short report the teacher receives on WhatsApp, with one thing to try next.

In Islamabad, a trained coach also sat in the lesson and filled a structured form on WhatsApp in two parts (the first half of the lesson, then the second), then checked Rumi's moments ("did you hear this happen?") and set a final level for each of 17 indicators. That record is in human/.

The rubrics are in rubrics/, in plain words. The prompts the models ran are in prompts/.

Files

README.md          this card
DICTIONARY.md      every file and every field
sessions.csv       one row per session (also sessions.parquet)
metadata.csv       the audio loader's index (file_name + the sessions.csv fields)
audio/             SES-####.mp3, mono 48 kbps
audio_original_codec/  the original upload, audio stream untouched (Opus .ogg, AAC .m4a or .mp3)
transcripts/       SES-####.txt as stored
tokens/            SES-####.json, word-level tokens with times and confidences
diarization/       SES-####.json, speaker segments
analysis/          SES-####.json, Rumi's scores, moments and report
human/             SES-####.json, the coach's record (Islamabad)
lesson_plans/      SES-####.txt, the linked lesson plan's text (18 Islamabad sessions)
rubrics/           one file per rubric
prompts/           the production prompts (start with INDEX.md)
MANIFEST.csv       sha256 and size of every file
totals.json        counts and hours

Where to start

Open sessions.csv. One row is one lesson; session_code names every other file of that lesson.
Pick a session, press play on audio/<session_code>.mp3, and read transcripts/<session_code>.txt beside it.
Open analysis/<session_code>.json for what Rumi concluded and told the teacher, and the matching rubric in rubrics/. For an Islamabad session, open human/<session_code>.json for the coach's record.

import pandas as pd, json
s = pd.read_csv("sessions.csv")
row = s.iloc[0]
analysis = json.load(open(row.analysis_file))

With the Hugging Face datasets library you can also load the audio and the sessions.csv fields together. Audio decoding needs torchcodec and torch as well as datasets (pip install datasets torchcodec torch):

from datasets import load_dataset
ds = load_dataset("<the dataset id in your invitation>", split="train")   # or the absolute path of your downloaded folder (the path snapshot_download returned)
ds[0]["audio"].get_all_samples()   # decodes one recording (48 kHz)

Without torchcodec, load_dataset stops with an error asking for it; pandas and the files work without it. The analysis/ files have two shapes, one for Islamabad and one for the other programmes (see DICTIONARY.md).

Words used here

Rubric (also called a framework): the checklist of teaching practices a lesson is judged on. Each programme uses its own: FICO-17 (Islamabad), HOTS and PECTAA (Pakistan), MEWAKA (Tanzania), the Palestine framework, TPAD (Kenya), with OECD, COT and FICO 1.0 on a few sessions. All are explained in rubrics/.
Indicator: one item on a rubric, scored on its own. Level: its grade, usually 1 to 4.
Transcript / tokens: the text the speech-to-text engine heard; tokens are its words with the time of each.
Speaker labels / diarization: splitting the recording into turns by speaker and labelling each turn Teacher or Student. Diarization is the field's word for it.
Moments: in Islamabad, short quotes from the transcript, each with its time, that Rumi picked as evidence for an indicator.
Sealed: the coach has finished both parts of the form and can no longer change them.

How it was prepared

Sessions were drawn from production by a fixed, seeded selection from every programme where Rumi coaches teachers (Yemen has had no sessions, so there is none here). One Kenya session was scored on FICO 1.0, so FICO sessions number 51 in totals.json. In Islamabad, sessions with a coach's form were chosen so that some teachers appear more than once.
Teacher, coach and session codes are new and made only for this dataset. Schools, districts, phone numbers and production ids are not included anywhere in the metadata.
Text and audio are as recorded and as stored. Children and teachers can be heard, and sometimes named, in the lessons, and names spoken in class appear in the transcripts. This is why access is private and the terms below apply.
One change to the stored text: where the system wrote a teacher's or coach's full name into a coach's brief or a report's WhatsApp caption, the name is replaced with their code (67 replacements: 46 briefs, 21 captions). First names, and everything in transcripts, tokens, speaker segments, moments and debriefs, are as stored.
Ten of the 18 lesson-plan texts (SES-1334, SES-2194, SES-4539, SES-4750, SES-5947, SES-6749, SES-8367, SES-9707, SES-9710, SES-9863) are unreadable extractions of Urdu PDFs, kept as stored; the coach's evidence_review.fidelity.moves[].text in human/<code>.json holds each plan's steps. SES-9710's record has no evidence review; its plan is the same file as SES-4539's, whose record holds the steps.
Photos of classrooms, which coaches sometimes take, are not included, to protect the children.
The original audio files had their container tags (device details) removed; the audio itself is untouched.
Some sessions were held back by us and are not in this dataset.
Teachers had other sessions that are not in this dataset; reports may refer to them.

Terms of use

Use this data only for the Taleemabad case study you were invited to. Do not share it or any part of it.
You may use it with any AI tool or provider you choose.
Delete every copy within 14 days of your call with us, including anything you uploaded to any service.
Do not try to identify any teacher, child, coach or school.

Contact

Reply to the email that invited you.

Downloads last month: 267

Total file size:

2.51 GB