isiZulu Synthetic Multi-Agent Conversations
Synthetic, knowledge-grounded, two-speaker conversations written entirely in isiZulu (zu).
Two LLM agents with sampled personas talk about a knowledge card (a source passage), and a separate
LLM judge scores each conversation. Only judge-accepted conversations are included here.
from datasets import load_dataset
ds = load_dataset("israel/zulu") # splits: train / validation / test
ds["train"][0]["conversation"] # [{"content": ..., "role": "user"}, {"content": ..., "role": "assistant"}, ...]
Splits
| split | conversations | share | knowledge cards | min judge mean | avg judge mean |
|---|---|---|---|---|---|
train |
2,438 | 69.2% | 491 | 5.50 | 6.19 |
validation |
378 | 10.7% | 80 | 5.50 | 6.13 |
test |
709 | 20.1% | 210 | 6.00 | 6.74 |
- No knowledge card is shared between splits. Splits are made over whole source passages, so validation and test measure generalization to unseen passages, not only unseen conversations.
- There is no textbook/gazette material for Zulu, so the test split is weighted toward the regionally valid source: native isiZulu Wikipedia. Each source sends a different share of its conversations to test (target: wikipedia 40%, sciq 15%, gsm8k 15%, bitext_cs 10%), while every source, including wikipedia, also keeps cards in train and validation. Test cards are picked best average judge score first and only conversations with judge
mean_score >= 6.0are kept. The lower-scored conversations of those cards are discarded rather than moved to train. - Test mix:
bitext_cs11%,gsm8k23%,sciq21%,wikipedia45%. - Train / validation: the remaining cards, assigned per source (~10% of each source to validation).
Per-source counts
| grounding_source | description | train | validation | test | total | test share |
|---|---|---|---|---|---|---|
bitext_cs |
Bitext customer-support dataset - English | 640 | 80 | 81 | 801 | 10% |
gsm8k |
openai/gsm8k math word problems - English | 764 | 110 | 162 | 1036 | 16% |
sciq |
allenai/sciq science questions - English | 714 | 107 | 148 | 969 | 15% |
wikipedia |
Native isiZulu Wikipedia (zu.wikipedia.org) - many South African people, places and culture | 320 | 81 | 318 | 719 | 44% |
Filtering (judge output)
Starting from 6,185 generated conversations, 3,525 are released. A conversation is kept only if:
- The LLM judge accepted it. The judge result must be valid (no error, all 4 scores in 1-7) and meet
the generation pipeline's acceptance rule: mean of the 4 scores >= 5.5, every score >= 4,
and
accuracy_score>= 5. - It passed the automatic QC (
quality.passed): correct turn count, alternating speakers, no empty turns, minimum length, no leaked speaker labels. - No turn is empty, and it is not an exact duplicate of another conversation.
- Test only: judge
mean_score >= 6.0.
Removed by steps 1-3: judge rejected or invalid: 2,526, qc failed: 15. Lower-scored conversations of held-out test cards were also removed.
The judge scores each conversation on a 1-7 Likert scale for instruction following, naturalness,
coherence and accuracy. Accuracy is checked against the knowledge card. Scores and short English
rationales are in the judge column.
Columns
| column | description |
|---|---|
conversation |
Chat format, same layout as israel/WikiConversations: list of {content, role}. Agent A becomes user, Agent B becomes assistant; always starts with user and alternates. |
turns |
Original turns: turn, speaker (Agent A / Agent B), persona_id, text. |
conversation_id, created_at |
ID and generation timestamp. The 8-char IDs are not guaranteed unique; use (conversation_id, created_at). |
language |
Conversation language (zu). |
grounding_source, grounding_language |
Source of the knowledge card and the card's language. |
knowledge_card |
card_id, title, topic, domain, difficulty, source_url, text (the grounding passage). |
persona_a, persona_b |
Role, register, knowledge level, personality, domain, communication style, code-switching level. |
goal, question_format |
Shared conversation goal; question type in the conversation (open, multiple_choice, true_false, fill_blank). |
model |
Generator LLM. |
quality |
Automatic QC flags. |
judge |
judge_model, the 4 scores with rationales, mean_score, accepted. |
Generation details
- Turns per conversation:
10(3,525) - Grounding language:
en(2,806),zu(719) - Generator models:
anthropic/claude-fable-5(2,394),google/gemini-3.7-flash(1,128),openai/gpt-5.6-terra(3) - Judge models (a different model from the generator where possible):
google/gemini-3.7-flash(1,582),openai/gpt-5.6-sol(999),anthropic/claude-fable-5(880),openai/gpt-5.6-terra(64) - Goals:
solve a problem(771),explain(745),clarify(744),learn(707),challenge(558) - Question formats:
open(2,831),multiple_choice(343),true_false(178),fill_blank(173) - Generated between 2026-09-12 and 2026-10-22.
Example (first two turns of a test conversation):
[
{
"content": "Sawubona mngani wami! Bengifunda nomfana wami incwadi yakhe ye-science izolo ebusuku, kwavela into engididayo — kuthiwa izinhlanganisela ezifana ne-sodium chloride, usawoti nje lo esiwusebenzisa ekhishini, zakha izakhiwo ezithile ezinohlelo oluphindaphindayo. Manje ngiyazibuza, ingabe uyakwazi ukuthi lezo zakhiwo zibizwa ngani, noma ngabe lokhu kuyiqiniso ngempela?",
"role": "user"
},
{
"content": "Haibo, awazi i-crystals? Yebo kuyiqiniso — i-sodium chloride yakha i-crystal, okuyi-rigid framework yama-ions amaningi abotshelwe ndawonye nge-repeating pattern. Kodwa ake ngikubuze: uyazi yini ukuthi ama-ions ayini, noma nikuvele nifundile nje nomfana ngaphandle kokuqonda ukuthi kwenzekani ngaphakathi?",
"role": "assistant"
}
]
Limitations
- The data is fully synthetic. It is LLM-generated and LLM-judged with no human review. Judge acceptance is a quality filter, not a guarantee of fluency or factual accuracy in a lower-resource language.
- Conversations grounded in English cards (
grounding_language: en) are model-written isiZulu and may contain code-switching (some personas ask for it). - The test set deliberately over-represents some sources (see above), so its source mix differs from train. Report per-source metrics where it matters.
Reproducing
Splits are built with split_and_push_zulu.py (seed 42). This card is generated with make_dataset_card.py.
- Downloads last month
- 85