Kinyarwanda Synthetic Multi-Agent Conversations

Synthetic, knowledge-grounded, two-speaker conversations written entirely in Kinyarwanda (rw). 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/kin")      # 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	8,420	69.5%	1406	5.50	6.26
validation	1,229	10.1%	208	5.50	6.23
test	2,471	20.4%	523	6.00	6.72

• 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.
• The test split contains only educational (rw_textbooks) and gazette (rw_gazette) material: Rwanda-specific, in-language sources that are the most valuable for evaluation. Whole knowledge cards are held out (best average judge score first, alternating between the two sources) and only conversations with judge mean_score >= 6.0 are kept. The lower-scored conversations of those held-out cards are discarded rather than moved to train.
• Test mix: rw_gazette 51%, rw_textbooks 49%.
• 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	1073	159	0	1232	0%
gsm8k	openai/gsm8k math word problems - English	1208	179	0	1387	0%
rw_gazette	Official Gazette of Rwanda 2012-2020 (laws, ministerial orders, appointments) - Kinyarwanda parts	1498	215	1253	2966	42%
rw_mono	mbazaNLP/kinyarwanda_monolingual_v01.1 (news, government, religious, legal text)	1068	155	0	1223	0%
rw_textbooks	REB Kinyarwanda textbooks, S1-S6 student books and teacher guides (O/A-Level)	1354	197	1218	2769	44%
sciq	allenai/sciq science questions - English	1062	155	0	1217	0%
wikipedia	Native Kinyarwanda Wikipedia (rw.wikipedia.org)	1157	169	0	1326	0%

Filtering (judge output)

Starting from 20,000 generated conversations, 12,120 are released. A conversation is kept only if:

1. 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.
2. It passed the automatic QC (quality.passed): correct turn count, alternating speakers, no empty turns, minimum length, no leaked speaker labels.
3. No turn is empty, and it is not an exact duplicate of another conversation.
4. Test only: judge mean_score >= 6.0.

Removed by steps 1-3: judge rejected or invalid: 7,620, qc failed: 25. 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 (rw).
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 (12,120)
• Grounding language: rw (8,284), en (3,836)
• Generator models: openai/gpt-5.6-sol (5,817), anthropic/claude-fable-5 (4,733), google/gemini-3.7-flash (1,570)
• Judge models (a different model from the generator where possible): google/gemini-3.7-flash (6,196), anthropic/claude-fable-5 (4,132), openai/gpt-5.6-sol (1,790), openai/gpt-5.6-terra (2)
• Goals: learn (2,592), explain (2,421), solve a problem (2,417), clarify (2,378), challenge (2,312)
• Question formats: open (7,478), multiple_choice (2,300), true_false (1,289), fill_blank (1,053)
• Generated between 2026-09-15 and 2026-10-21.

Example (first two turns of a test conversation):

[
  {
    "content": "Uyu munsi turasesengura imigani migufi, cyane cyane ngo «Amatwi yuje amata ntiyumva.» Urumva uyu mugani usobanura iki, kandi wawuca ryari ubonye umuntu wateye imbere agahindura indoro n’ingendo, agasuzugura abo bahoranye mu bukene?",
    "role": "user"
  },
  {
    "content": "Njye numva uvuga ko hari umuntu ukize cyangwa wagize icyo yunguka, maze akirengagiza agasuzugura abo bahoranye mu bukene, nk’aho atakibumva. Nawuca igihe mbonye uwateye imbere agahindura imyitwarire n’ingendo kubera kwibagirwa ko iby’isi ari intizo; ariko se, mwarimu, buri muntu uhinduye indoro amaze gukira twahita tumuciraho uyu mugani, cyangwa twabanza kureba niba koko asuzugura abandi?",
    "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 Kinyarwanda 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_kin.py (seed 42). This card is generated with make_dataset_card.py.