oddadmix/arabic-rag-chat-grpo-5K

Arabic multi-turn RAG conversations — GRPO pool (5,259 conversations)

The reinforcement-learning half of `oddadmix/arabic-rag-chat-30K`: same generator, same validator, same schema, disjoint companies. It exists so GRPO explores fresh knowledge bases instead of taking a second pass over material the SFT already memorised.

Company-disjointness is exact and verified: this pool shares zero company_id values with either the train or test split of the 30K set, and those two share none with each other.

Composition

55% of conversations carry a customer account record.

Fields

Identical to the 30K set — see that card for the full field list and the validation rules, which are the same here. In short: chunks are the numbered passages, account is the out-of-band customer record, and each turn carries gold_chunk_ids plus per-part key_facts that are validated to occur verbatim in the cited passage.

There is one split (data/all.jsonl) because this pool is never evaluated on — evaluation happens on the 30K set's test split, and the per-row artifacts are published separately as `oddadmix/arabic-rag-chat-8k-eval`.

How the reference recipe consumes it

One prompt per conversation, one sampled round — not every round. Rounds of the same conversation share a knowledge base and most of a prompt, so training on all of them would spend the rollout budget re-reading the same passages and make the effective batch far more correlated than its size suggests.

Prompts over 4,096 tokens are dropped before priority selection, not after: priority correlates hard with length, so filtering second would take the top n and then delete most of them. On the reference 50M run that left 3,046 usable prompts of 5,259 conversations — 2,213 dropped on length, and of what remained 1,270 multi-gold, 725 multi-part, 202 refusal.

Selection is failure-weighted toward multi-part and multi-gold turns, with numeric mixing hazard as the tie-break. Refusal rows are held at a fixed 15% share so the policy is not pushed toward declining more often, which is the cheapest way to game two of the seven rewards.

The rewards are programmatic — no judge in the training loop: number_grounding (1.0), refusal_reward (0.6), chrf_reward (0.8), sanity_reward (0.3), part_coverage (0.8), aggregation_recall (0.6), gold_coverage (1.0). part_coverage and aggregation_recall are what the per-part key_facts labels in this schema exist to make computable.

Caveats

All content is fictional and machine-generated; internally consistent per company, but not real. The teacher was instructed never to calculate — answers quote figures rather than combining them — so this set does not teach arithmetic, and number_grounding is correspondingly a strict verbatim string test rather than a numerical-reasoning check.

Measured GRPO gains from this pool were concentrated in per-passage coverage, not in aggregation: across three reward revisions part-coverage moved 50.1% → 58.5% while aggregation-recall stayed flat at 53.0% → 53.7%. If you are using this pool to attack multi-passage aggregation specifically, that is the honest prior — see the eval dataset for the full table.

© KAND CA 2026 — PROJECT NAWAH.