huggingface/transformers · error · TypeError

Training input {text_a} is not a string

Error message

Training input {text_a} is not a string

What it means

While building XNLI training examples from the multilingual TSV, the processor asserts that column 0 of each row (text_a) is a str. The CSV reader normally yields strings, so this TypeError signals a malformed row: wrong delimiter, a file without a header as expected, QUOTE_NONE artifacts, or a row whose columns are not plain text. The check exists to fail loudly on corrupt data instead of producing broken InputExamples.

Source

Thrown at src/transformers/data/processors/xnli.py:49

    def __init__(self, language, train_language=None):
        self.language = language
        self.train_language = train_language

    def get_train_examples(self, data_dir):
        """See base class."""
        lg = self.language if self.train_language is None else self.train_language
        lines = self._read_tsv(os.path.join(data_dir, f"XNLI-MT-1.0/multinli/multinli.train.{lg}.tsv"))
        examples = []
        for i, line in enumerate(lines):
            if i == 0:
                continue
            guid = f"train-{i}"
            text_a = line[0]
            text_b = line[1]
            label = "contradiction" if line[2] == "contradictory" else line[2]
            if not isinstance(text_a, str):
                raise TypeError(f"Training input {text_a} is not a string")
            if not isinstance(text_b, str):
                raise TypeError(f"Training input {text_b} is not a string")
            if not isinstance(label, str):
                raise TypeError(f"Training label {label} is not a string")
            examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
        return examples

    def get_test_examples(self, data_dir):
        """See base class."""
        lines = self._read_tsv(os.path.join(data_dir, "XNLI-1.0/xnli.test.tsv"))
        examples = []
        for i, line in enumerate(lines):
            if i == 0:
                continue
            language = line[0]
            if language != self.language:
                continue
            guid = f"test-{i}"

View on GitHub (pinned to a597f97485)

Solutions

  1. Re-download the XNLI-MT-1.0 archive and verify multinli.train.{language}.tsv opens as tab-separated text with a header row.
  2. Sanity-parse the file yourself with csv.reader(f, delimiter='\t', quoting=csv.QUOTE_NONE) and inspect the offending row index.
  3. Confirm data_dir contains XNLI-MT-1.0/multinli/multinli.train.{lg}.tsv exactly as the processor expects.

Example fix

# before
examples = processor.get_train_examples(data_dir)  # corrupt TSV

# after: validate the file first
import csv
with open(f"{data_dir}/XNLI-MT-1.0/multinli/multinli.train.en.tsv", encoding="utf-8") as f:
    rows = list(csv.reader(f, delimiter="\t", quoting=csv.QUOTE_NONE))
assert all(isinstance(r[0], str) for r in rows[1:])
examples = processor.get_train_examples(data_dir)
Defensive patterns

Strategy: validation

Validate before calling

import csv
path = f"{data_dir}/XNLI-MT-1.0/multinli/multinli.train.{lg}.tsv"
with open(path, encoding="utf-8") as f:
    rows = list(csv.reader(f, delimiter="\t", quoting=csv.QUOTE_NONE))[1:]
assert rows and all(isinstance(r[0], str) and r[0] for r in rows), "malformed text_a column"

Try / catch

try:
    examples = processor.get_train_examples(data_dir)
except TypeError as e:
    if "not a string" in str(e):
        raise ValueError(f"Corrupt XNLI train TSV under {data_dir}; re-download XNLI-MT-1.0") from e
    raise

Prevention

When it happens

Trigger: get_train_examples(data_dir) on a multinli.train.{lg}.tsv that is corrupted, uses a different delimiter/quoting, or where line[0] is empty/None-like; pointing data_dir at the wrong directory so a different file layout is read.

Common situations: Downloading XNLI-MT-1.0 manually and getting HTML or partial content; locale/encoding issues producing bytes; editing the TSV in a spreadsheet that changed quoting.

Related errors

AI-assisted analysis of huggingface/transformers@a597f97485 (2026-08-14). Data as JSON: /api/errors/c13335ea01f6fe86. Report an issue: GitHub.