Update README.md
Browse files
README.md
CHANGED
@@ -84,54 +84,61 @@ names_to_remove = [
|
|
84 |
"department establishment",
|
85 |
]
|
86 |
|
87 |
-
def extract_intents_data(events_dataset: HFDataset) ->
|
88 |
"""Extract intent names and assign ids to them."""
|
89 |
intent_names = sorted({name for intents in events_dataset["train"]["all_labels"] for name in intents})
|
90 |
for n in names_to_remove:
|
91 |
intent_names.remove(n)
|
92 |
-
|
93 |
-
intents_data = [Intent(id=i,name=name) for i, name in enumerate(intent_names)]
|
94 |
-
return intents_data, name_to_id
|
95 |
|
96 |
|
97 |
-
def converting_mapping(example: dict,
|
98 |
-
"""Extract utterance and label and drop the rest."""
|
99 |
-
|
100 |
"utterance": example["content"],
|
101 |
"label": [
|
102 |
-
|
103 |
-
]
|
104 |
}
|
|
|
|
|
|
|
105 |
|
106 |
|
107 |
-
def convert_events(events_split: HFDataset,
|
108 |
"""Convert one split into desired format."""
|
109 |
events_split = events_split.map(
|
110 |
converting_mapping, remove_columns=events_split.features.keys(),
|
111 |
-
fn_kwargs={"
|
112 |
)
|
113 |
|
114 |
-
|
115 |
-
oos_samples = [] # actually this dataset doesn't contain oos_samples so this will stay empty
|
116 |
for sample in events_split.to_list():
|
117 |
if sample["utterance"] is None:
|
118 |
continue
|
119 |
-
|
120 |
-
sample.pop("label")
|
121 |
-
oos_samples.append(sample)
|
122 |
-
else:
|
123 |
-
in_domain_samples.append(sample)
|
124 |
|
125 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
126 |
|
127 |
if __name__ == "__main__":
|
128 |
-
#
|
|
|
129 |
events_dataset = load_dataset("knowledgator/events_classification_biotech", trust_remote_code=True)
|
130 |
|
131 |
-
intents_data
|
132 |
|
133 |
-
train_samples = convert_events(events_dataset["train"],
|
134 |
-
test_samples = convert_events(events_dataset["test"],
|
135 |
|
136 |
events_converted = Dataset.from_dict(
|
137 |
{"train": train_samples, "test": test_samples, "intents": intents_data}
|
|
|
84 |
"department establishment",
|
85 |
]
|
86 |
|
87 |
+
def extract_intents_data(events_dataset: HFDataset) -> list[Intent]:
|
88 |
"""Extract intent names and assign ids to them."""
|
89 |
intent_names = sorted({name for intents in events_dataset["train"]["all_labels"] for name in intents})
|
90 |
for n in names_to_remove:
|
91 |
intent_names.remove(n)
|
92 |
+
return [Intent(id=i,name=name) for i, name in enumerate(intent_names)]
|
|
|
|
|
93 |
|
94 |
|
95 |
+
def converting_mapping(example: dict, intents_data: list[Intent]) -> dict[str, str | list[int] | None]:
|
96 |
+
"""Extract utterance and OHE label and drop the rest."""
|
97 |
+
res = {
|
98 |
"utterance": example["content"],
|
99 |
"label": [
|
100 |
+
int(intent.name in example["all_labels"]) for intent in intents_data
|
101 |
+
]
|
102 |
}
|
103 |
+
if sum(res["label"]) == 0:
|
104 |
+
res["label"] = None
|
105 |
+
return res
|
106 |
|
107 |
|
108 |
+
def convert_events(events_split: HFDataset, intents_data: dict[str, int]) -> list[Sample]:
|
109 |
"""Convert one split into desired format."""
|
110 |
events_split = events_split.map(
|
111 |
converting_mapping, remove_columns=events_split.features.keys(),
|
112 |
+
fn_kwargs={"intents_data": intents_data}
|
113 |
)
|
114 |
|
115 |
+
samples = []
|
|
|
116 |
for sample in events_split.to_list():
|
117 |
if sample["utterance"] is None:
|
118 |
continue
|
119 |
+
samples.append(sample)
|
|
|
|
|
|
|
|
|
120 |
|
121 |
+
mask = [sample["label"] is None for sample in samples]
|
122 |
+
n_oos_samples = sum(mask)
|
123 |
+
n_in_domain_samples = len(samples) - n_oos_samples
|
124 |
+
|
125 |
+
print(f"{n_oos_samples=}")
|
126 |
+
print(f"{n_in_domain_samples=}\n")
|
127 |
+
|
128 |
+
# actually there are too few oos samples to include them, so filter out
|
129 |
+
samples = list(filter(lambda sample: sample["label"] is not None, samples))
|
130 |
+
|
131 |
+
return [Sample(**sample) for sample in samples]
|
132 |
|
133 |
if __name__ == "__main__":
|
134 |
+
# `load_dataset` might not work
|
135 |
+
# fix is here: https://github.com/huggingface/datasets/issues/7248
|
136 |
events_dataset = load_dataset("knowledgator/events_classification_biotech", trust_remote_code=True)
|
137 |
|
138 |
+
intents_data = extract_intents_data(events_dataset)
|
139 |
|
140 |
+
train_samples = convert_events(events_dataset["train"], intents_data)
|
141 |
+
test_samples = convert_events(events_dataset["test"], intents_data)
|
142 |
|
143 |
events_converted = Dataset.from_dict(
|
144 |
{"train": train_samples, "test": test_samples, "intents": intents_data}
|