Orsa NLU intent model v8
Multi-label (sigmoid + per-intent thresholds in thresholds.json) XLM-R intent classifier for the Orsa voice-ordering speaker, ru / kk / en.
20 intents: the v6 taxonomy with repeat_order replaced by my_order_info — the guest asks what is in their current order
(«что у меня в заказе», «повторите мой заказ», «тапсырысымда не бар», "what did I order", «а суп я уже заказывал?»).
Promoted to main on 2026-10-20. Rollback: branch v6 (previous model, 20 intents with repeat_order + farewell), v5, v4, v3-legacy.
The orchestrator must know my_order_info before deploying (its ParseIntent maps an unknown label to unknown); this model never emits repeat_order.
What changed vs v6
- There is no order history, so re-order requests can't be served: 357 history
repeat_orderrows («как обычно», «как в прошлый раз», "same as last time") were dropped; 5 plain read-backs («повторите мой заказ», «тапсырысты қайталаңыз») moved tomy_order_info(nlu_finetune_orsa/taxonomy_v8.py). - ~250 rows from
data/seeds_v8_{ru,kk,en}.json: order read-back phrasings, "did I order X" templates, contrasts («что входит в плов» →ingredient_info, «отмените мой заказ» →cancel_order) and compounds (my_order_info + request_bill / order / call_waiter). - Order totals («сколько стоит мой заказ») stay
price/request_bill.
Results (top-1 single, exact set for compounds)
| v6 | v8 | |
|---|---|---|
eval_v8_order_info (frozen) |
17/61 | 61/61 |
eval_v4 |
222/237 | 222/237 |
eval_v4b |
170/180 | 171/180 |
eval_v5_holdout |
158/159 | 156/159 |
eval_v6_farewell |
49/51 | 50/51 |
eval_v7_wish |
30/33 | 30/33 |
(repeat_order items excluded from the old sets.) Kazakh v8 seeds are not yet native-reviewed.
- Downloads last month
- 62
Model tree for Damilya/nlu-restaurant-multilingual
Base model
FacebookAI/xlm-roberta-base