bin-sorting-intents

Synthetic voice-command sentences for a robot cell with two destination boxes (big and white) and two kinds of pieces (with or without a bolt, good or bad), in Spanish, English and Basque. Each sentence is labelled with five classification answers. It trains IkerAzurmendi/bin-intent-gliner2.

{"input": "Clasifica las piezas malas en la caja blanca",
 "output": {"classifications": [
   {"task": "intent",  "labels": ["move", "sort", "pick", "both", "reply", "ask", "stop", "cancel", "repeat", "yes", "no", "hello", "chat", "refuse"], "true_label": ["sort"]},
   {"task": "feature", "labels": ["bolt", "no bolt", "any"], "true_label": ["any"]},
   {"task": "quality", "labels": ["good", "bad", "any"], "true_label": ["bad"]},
   {"task": "box",     "labels": ["big", "white", "other", "none"], "true_label": ["white"]},
   {"task": "count",   "labels": ["one", "two", "three", "all", "none"], "true_label": ["all"]}]},
 "meta": {"lang": "es", "tier": "train", "intent": "sort"}}

This is the format the GLiNER2 trainer reads. box is the big box or the white box; other is a box the cell does not have ("the small one", "the left one"); none is no destination.

Splits

file	sentences	what it is
train.jsonl	60,000 (20,000 per language)	built from the training word lists
eval.jsonl	2,997	held out of train, same words
calib.jsonl	2,400	built with a second set of words that never occur in train; used to fit probability calibration
test_unseen.jsonl	2,997	built with a third set of words that occur in neither; measures generalisation
test_real.txt	143	written by hand (lang | sentence | intent | feature | quality | box | count)

No sentence appears in two files (compared without case, accents or punctuation). stats.md has the counts per intent and language. 62% of the training sentences are orders (move, sort, pick, both); the rest are replies, questions, control words, chat and refusals (negations, destruction, things that are not pieces, past statements). About a quarter of the sentences lose their punctuation, 12% of Spanish and Basque ones lose accents, 5% get a typo and 2% a plausible speech-recognition mishearing; 7% of move/sort orders contain a self-correction ("a la grande, no, a la blanca": the last thing said wins).

generator/ holds the code that produced it (generate_dataset.py and the three grammars); the word lists are in the grammar_*.py files, split into train / calib / test tiers.

Limits

• All sentences are generated from grammars, not collected from people. The 143-sentence test_real.txt was written by the same author as the grammars. Neither says how real operators or a speech recogniser will phrase things.
• Basque was written without a native speaker and is a synthetic approximation of spoken Basque.
• The label space is specific to this cell (two boxes, bolt, good/bad); it will not transfer to another layout without new grammar words.