MohammadAlnajdawi/sheikhchat-transcript-corpus

SheikhChat — transcript corpus + candidate instruction pairs

Working dataset for fine-tuning a small model on the voice and format of a single Arabic religious speaker. It is a staging artifact, not a production training set.

Read this before training on it

The pair labels are not verified. The source transcripts have no speaker labels at all — every segment is [UNKNOWN], and the >> markers are ASR phrase boundaries (median segment ~30 characters), not speaker turns. 19 of 77 files (23% of the corpus) have no markers whatsoever.

Pairs were therefore produced by a heuristic, and audited by hand:

stage	method	outcome
v1	loose heuristic	3/12 genuine (~25%)
v2	tightened heuristic	6/10 genuine (~60%)
v3	LLM judge, categorical rubric	102/164 VALID (62.2%)
v3	judge hand-audited	11/12 correct (~92%)

The regex filters were abandoned: applied to all 164 pairs they removed only 14%, while a hand audit found ~40% defective. The defects are semantic — a caller's line sitting in the teacher's slot, stream housekeeping, ASR that reads fluent but means nothing — and no keyword list sees them.

Only pairs the judge labelled VALID are in train.jsonl. Judge verdicts survive in judge_labels.json (VALID / CALLER / ADMIN / INCOHERENT / OFF_TOPIC / UNCLEAR) so the filter is reproducible and arguable rather than a black box.

This is a pre-filter, not a certificate. ~8% judge error on 102 rows means roughly eight wrong pairs remain, and the judge cannot see everything a human would. Use data/pairs_review.csv for the final accept/reject pass.

What is in here

data/corpus.jsonl         cleaned transcript corpus, one document per source
data/pairs_review.csv     the JUDGE-VALIDATED pairs + empty review columns
judge_labels.json         every judge verdict, for every candidate pair
train.jsonl               validated pairs + authored refusals (chat format)
val.jsonl
test.jsonl
refusals.jsonl            the authored refusal/scope examples, on their own
merge_stats.json          persona vs refusal counts per split

Two kinds of row, and only one of them is transcribed

Every pair carries meta.kind:

kind	origin	trust
persona	extracted from the transcripts by heuristic	~60% (see above)
refusal	authored — written for this dataset, not transcribed	synthetic

The refusals exist because a set built only from real answers teaches a model that every question deserves an answer. That is the wrong behaviour for a religious assistant: it will rule on someone's divorce or invent a hadith rather than decline. The set covers six failure modes — personal fatwa, out-of-scope, identity/escalation, crisis, live data, and insults — and the canonical replies are copied from the running application (jev_escalate_reply, jev_out_of_scope_reply, BLOCKED_REPLY) so training matches serving rather than drifting from it.

Ratio in train.jsonl: 24 of 175 rows (13.7%). Val and test each hold three refusals, one per behaviour class where possible, so "did it learn to decline?" is measurable instead of assumed.

Schema

corpus.jsonl

{"id": "<video id>", "text": "<cleaned transcript>",
 "chars": 12345, "source_url": "..."}

train/val/test.jsonl — chat format, ready for TRL / Unsloth:

{"messages": [{"role": "system", "content": "..."},
              {"role": "user", "content": "<caller question>"},
              {"role": "assistant", "content": "<the speaker's answer>"}],
 "meta": {"source": "<video id>", "confidence": 0.9, "kind": "persona"}}

Cleaning applied

Removed: N seconds/minutes ASR timestamp artifacts (present in ~14% of indexed chunks before cleaning), [UNKNOWN] / [تنحنُح]-style tags, Latin-script ASR drift (the model wandering into English/German), tatweel, duplicated whitespace.

Normalization of diacritics was deliberately not baked into the stored text. That belongs at query time — baking it in trades readability for a matching gain you can get for free at search time.

Splits

Split by source video, never per example: passages from one multi-hour stream repeat each other, and a per-example split leaks the test set. Sizes are small (151 / 4 / 9) because source files, not pairs, are the unit.

Known limitations

• No speaker attribution. The central defect. Proper (question → teacher answer) extraction needs audio diarization, not regexes over ASR text.
• Dialect. Egyptian colloquial throughout. Train on this and the model answers in colloquial — correct for persona, but decide it deliberately.
• ASR errors propagate. Transcripts are auto-captions: no punctuation (0.33% of characters), misrecognised names, dropped words.
• Refusals are authored, not observed. The 30 refusal rows are synthetic — written to encode the product's intended behaviour. They teach the decision to decline; they are not evidence of how the speaker actually phrases a refusal, because the corpus contains almost none.
• Crisis rows are minimal. Three examples. Enough to establish that a self-harm or abuse message must not be answered as a fatwa, not enough to be a safety guarantee. Do not deploy on the strength of this dataset alone.
• Facts do not belong here. Fine-tune for voice and format; keep doctrine in retrieval. A small model trained on facts states them confidently and wrongly.

Provenance and rights

Derived from publicly available video transcripts of a single speaker, assembled for a private research project. Rights in the underlying recordings are not cleared here. Treat as private research data; do not redistribute without confirming you have the right to.