oddadmix/arabic-rag-chat-30K

Arabic multi-turn RAG customer-support conversations (31,294 conversations)

Synthetic Modern Standard Arabic customer-support conversations for training small Arabic RAG assistants. The bulk was distilled from gemini-3.1-flash-lite via the Batch API; a first 2.4% came from unsloth/gemma-4-31B-it-NVFP4 on a local vLLM server before the run was moved off-GPU. Both teachers were given the same prompts and the same validator. Each row is one conversation of 1-5 rounds over one knowledge base, with an optional per-customer account record.

What this set is for

It exists to train three things a single-turn, single-passage corpus cannot:

1. Multi-context aggregation - 47% of answerable turns require two or more passages that are neither adjacent nor from the same section.
2. Multi-part coverage - 25% of turns ask 2-3 separate things in one message. Each part carries its own gold_chunk_ids and its own key_facts, so "did the reply cover every part" is a string test, not a judgement call.
3. Account-aware grounding - 56% of conversations carry a customer record (account number, plan, outstanding balance, due date, last payment, usage) supplied outside the numbered passages, which must be preferred over generic policy text.

8% of turns are unanswerable from their own passages, many of them mid-conversation: a model that has answered twice must still be able to decline.

Length buckets

Each company has ONE knowledge base; the three buckets are prefixes of it, stored round-robin across aspect sections so that even the shortest prefix spans every aspect rather than being several passages about pricing.

bucket	conversations	passages p50	tokens p50
short	10,464	6	1,070
mid	10,381	26	3,809
long	10,449	50	7,112

Roughly equal by conversation count, which is roughly 59% of tokens in the long bucket. Training on it directly puts most of the gradient on long rows; the reference training recipe corrects that with a short-weighted anneal epoch rather than by rebalancing the data.

Fields

id, company_id (rows sharing a value share a knowledge base), company, domain, country, bucket, chunks (list[str], the MSA passages), account (str or null - the customer record, rendered as one text block), hard_negative_ids (non-gold passages carrying a different figure of the same kind), turns, tone, n_chunks, n_rounds, n_tokens.

Each entry of turns has: question, answer, answerable, turn_type (single | multipart | followup | new-constraint | correction | unanswerable | partial), uses_account, gold_chunk_ids (0-based indices into chunks; empty when unanswerable), and parts - one entry per separate ask, each with ask, its own gold_chunk_ids, and key_facts.

Validation

Rows are dropped, never repaired. Every published turn satisfies: gold ids in range; every key_facts string occurs verbatim in that part's cited passages or in the account record and in the answer; every distinctive number (10 and above) in the answer occurs in its cited passages, the account record, or the question; unanswerable turns cite nothing and carry an explicit refusal; Arabic ratio above 0.6; no duplicate questions within a conversation; the whole conversation fits 7,800 student tokens. A turn that fails truncates its conversation at that point - earlier rounds are kept.

Splits

split	conversations	turns	companies
train	30,662	81,579	1,815
test	632	1,651	37

Split by company, so train and test never share a knowledge base.

Caveats

All content is fictional and machine-generated; facts are internally consistent per company but not real. Intended for training models to copy from provided context, not as a knowledge source. The teacher was instructed never to calculate - answers quote figures rather than combining them - so this set does not teach arithmetic. parts labels are teacher-supplied and only partially checkable: a wrong gold_chunk_ids whose key_facts happen to occur in the cited passage anyway will pass validation.

© KAND CA 2026 - PROJECT NAWAH.