Data Structures

Eevee works with CSV label dataframes with items as per these definitions. Since label dataframes have id for referring back to the data, we just focus on labels in this tool. These labels can be true labels of various kind or coming from predictions of different models.

This page documents a few general notes about the label representation. Specific details are in the pages for different kinds of metrics here.

Serialization

Each row in the label dataframe CSV is of one of the types defined here. In cases where the field type are not primitives, we serialize items in JSON. In Python, this looks like the following:

import pandas as pd

# Assuming each item in `items` is a list of entities
rows = [{"id": i, "entities": json.dumps(it)} for i, it in enumerate(items)]

pd.DataFrame(rows).to_csv("./predictions.csv", index=False)

The following is how correctly serialized structure looks like in a labels CSV:

"[[{""am_score"": -278.4794, ""confidence"": 0.9739978, ""lm_score"": 13.827044, ""transcript"": ""no""}]]"

If you skip JSON dumping, tools like pandas might still serialize like following:

"[[{'am_score': -278.4794, 'confidence': 0.9739978, 'lm_score': 13.827044, 'transcript': 'no'}]]"

But this won’t be read back in eevee and you will get a JSONDecodeError

JSONDecodeError: Expecting property name enclosed in double quotes: line 1 column 4 (char 3)