Looking for the format of the dataset:
Are there train and test spits on this dataset, and is there a format were the tags column is the label_ID's rather than the labels themselves? I'm having some trouble changing the dataset into a similar format as other NER datasets like: ncbi_disease etc
The format is created to be compatible with Huggingface AutoModelForTokenClassification. I didn't create any other format. You can find my example training notebook here:
https://colab.research.google.com/drive/1OzCY782KJSF0FBDS0d1CoMhfp3-RtJMV?usp=sharing
Ah thanks so much i'll play around with it a little.
from datasets import load_dataset
import numpy as np
#But then after that it's back to normal
try:
# Try to get the feature names directly
feature_names = MACCROBAT['train'].features['tags'].feature.names
except AttributeError:
# If 'feature.names' is not available, get the unique tags as before
all_tags = np.concatenate(MACCROBAT['train']['tags'])
unique_labels = np.unique(all_tags)
feature_names = sorted(unique_labels)
# Print sorted labels
print(MACCROBAT)
for i, name in enumerate(feature_names):
print(f"{i}: {name}")
MACCROBAT_original_labels = {
0: "B-Activity",
1: "B-Administration",
2: "B-Age",
3: "B-Area",
4: "B-Biological_attribute",
5: "B-Biological_structure",
6: "B-Clinical_event",
7: "B-Color",
8: "B-Coreference",
9: "B-Date",
10: "B-Detailed_description",
11: "B-Diagnostic_procedure",
12: "B-Disease_disorder",
13: "B-Distance",
14: "B-Dosage",
15: "B-Duration",
16: "B-Family_history",
17: "B-Frequency",
18: "B-Height",
19: "B-History",
20: "B-Lab_value",
21: "B-Mass",
22: "B-Medication",
23: "B-Nonbiological_location",
24: "B-Occupation",
25: "B-Other_entity",
26: "B-Other_event",
27: "B-Outcome",
28: "B-Personal_background",
29: "B-Qualitative_concept",
30: "B-Quantitative_concept",
31: "B-Severity",
32: "B-Sex",
33: "B-Shape",
34: "B-Sign_symptom",
35: "B-Subject",
36: "B-Texture",
37: "B-Therapeutic_procedure",
38: "B-Time",
39: "B-Volume",
40: "B-Weight",
41: "I-Activity",
42: "I-Administration",
43: "I-Age",
44: "I-Area",
45: "I-Biological_attribute",
46: "I-Biological_structure",
47: "I-Clinical_event",
48: "I-Color",
49: "I-Coreference",
50: "I-Date",
51: "I-Detailed_description",
52: "I-Diagnostic_procedure",
53: "I-Disease_disorder",
54: "I-Distance",
55: "I-Dosage",
56: "I-Duration",
57: "I-Family_history",
58: "I-Frequency",
59: "I-Height",
60: "I-History",
61: "I-Lab_value",
62: "I-Mass",
63: "I-Medication",
64: "I-Nonbiological_location",
65: "I-Occupation",
66: "I-Other_entity",
67: "I-Other_event",
68: "I-Outcome",
69: "I-Personal_background",
70: "I-Qualitative_concept",
71: "I-Quantitative_concept",
72: "I-Severity",
73: "I-Shape",
74: "I-Sign_symptom",
75: "I-Subject",
76: "I-Texture",
77: "I-Therapeutic_procedure",
78: "I-Time",
79: "I-Volume",
80: "I-Weight",
81: "O",
}
from sklearn.model_selection import train_test_split
# Assume that 'dataset' is a list of dictionaries with 'tokens' and 'tags'
tokens = [example['tokens'] for example in MACCROBAT]
tags = [example['tags'] for example in MACCROBAT]
# Split the dataset into training and testing sets
tokens_train, tokens_test, tags_train, tags_test = train_test_split(tokens, tags, test_size=0.2, random_state=42)
# Convert the split data back into the original format
train_data = [{'tokens': tok, 'tags': tag} for tok, tag in zip(tokens_train, tags_train)]
test_data = [{'tokens': tok, 'tags': tag} for tok, tag in zip(tokens_test, tags_test)]
# Create a dictionary with the split datasets
MACCROBAT = {'train': train_data, 'test': test_data}
This is kind of what I was trying to do, so I can make it a similar format with the others so I can combine them, but I think the problem is that the tokenizer is reading the str label, not the Label_ID. Idk if you have any quick fixes, but I appreciate the reply from earlier I'll check it out.
Thanks again
-Doug
Hi in the original dataset each sample is a text, but here its a list of tokens. How did you split the text into tokens and align the labels? Can you share the code that transformed the original data to this format?
Thanks