Chapter 04 - text-classification pipeline returns single dict instead of list

[HLZHarry]

When running the text-classification inference loop from the book, the output returned by the pipeline is a single dictionary instead of a list of dictionaries, causing a KeyError when trying to access output[0].

Code from the book

y_pred = []
for output in tqdm(pipe(KeyDataset(data["test"], "text")), total=len(data["test"])):
    negative_score = output[0]["score"]
    positive_score = output[2]["score"]
    assignment = np.argmax([negative_score, positive_score])
    y_pred.append(assignment)

Expected Output

[{"label": "negative", "score": 0.73}, {"label": "neutral", "score": 0.15}, {"label": "positive", "score": 0.12}]

Actual Output

{"label": "positive", "score": 0.95}

Error

KeyError: 0

My Fix

Replace return_all_scores=True with top_k=None:

pipe = pipeline(
    model=model_path,
    tokenizer=model_path,
    top_k=None,   # replaces deprecated return_all_scores=True
    device="cuda:0"
)

[HLZHarry]

Fix

Instead of accessing by index, convert the output to a dictionary and access by label name:

y_pred = []
for output in tqdm(pipe(KeyDataset(data["test"], "text")), total=len(data["test"])):
    scores = {d["label"]: d["score"] for d in output}  # convert to dict {label: score}
    negative_score = scores["negative"]
    positive_score = scores["positive"]
    assignment = np.argmax([negative_score, positive_score])
    y_pred.append(assignment)

[MaartenGr]

Could you share in which environment you are working? Specifically, we intended to fix package versions so that API changes wouldn't be an issue.

[HLZHarry]

Environment Details

• Platform: Google Colab
• GPU: T4 (Colab free tier)
• Python version: 3.12.12 (main, Oct 10 2025, 08:52:57) [GCC 11.4.0]
• Transformers version: 5.0.0
• PyTorch version: 2.10.0+cu128

I installed the latest versions of the libraries directly in Colab.

[MaartenGr]

That's good to know that you used the latest versions of the libraries! Ideally, we want to pin the versions to what's in requirements.txt as a way to make everything reproducible. Some libraries have a tendency of changing their API frequently, which would require continuously updating all code examples. However, we left the versioning open to see how long packages, for each chapter, would continue to have the same API. It seems that it has changed for what we wanted to achieve in Chapter 4, so if you pin the versions, that might work!

If you want to test that, then that would be great! If it works, I can go ahead and pin the versions in that notebook.