docs: replace deprecated financial_phrasebank dataset in IA3 tutorial

[dhruvildarji]

Summary

Fixes #2998

The financial_phrasebank dataset fails to load with recent versions of the datasets library because it relied on a deprecated loading script format. This PR replaces it with zeroshot/twitter-financial-news-sentiment, a financial sentiment dataset that uses the modern parquet format.

Changes

• docs/source/task_guides/ia3.md: updated dataset reference, description, and text_column from "sentence" to "text" to match the new dataset's column name
• examples/conditional_generation/peft_adalora_seq2seq.py: same dataset and column name updates

The replacement dataset (zeroshot/twitter-financial-news-sentiment) was tested and confirmed working by @maerory in the issue thread.

Test plan

• Updated dataset loads successfully with load_dataset("zeroshot/twitter-financial-news-sentiment")
• Column name updated from sentence to text throughout affected files
• Label names still auto-detected from features["label"].names (no hardcoded changes needed)

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[dhruvildarji]

Thanks for the review @BenjaminBossan! I've added the conditional generation notebook changes to this PR as requested. The changes from #3059 are now included here — the fix covers all instances: IA3 tutorial docs, peft_adalora_seq2seq.py, and all 4 conditional generation notebooks (IA3, LoRA, prompt tuning, prefix tuning).

[BenjaminBossan]

Thanks for replacing the outdated dataset with a newer one.

I found one more instances where it appears: examples/int8_training/Finetune_flan_t5_large_bnb_peft.ipynb. Could you please fix that notebook too?

I also saw that these notebooks define:

checkpoint_name = "financial_sentiment_analysis_<peft-method>.pt"

With the changed dataset, the name doesn't really fit anymore. However, the variable isn't used at all, so in a sense it doesn't matter. But if you're up to it, it would be great to remove it completely and avoid possible confusion.

[BenjaminBossan]

Thanks for the update @dhruvildarji. I tried running one of the notebooks (peft_lora_seq2seq) and encountered some issues. First of all, the line:

classes = dataset["train"].features["label"].names

didn't work but I could replace it with:

classes = dataset["train"]["label"]

The next issue came from this line in preprocess_function:

labels = tokenizer(targets, max_length=3, padding="max_length", truncation=True, return_tensors="pt")

The issue here is that the targets are not strings, so they cannot be tokenized. Instead, they are ints, which encode the sentiment classes. Now we could just use those directly as the target without passing them through the tokenizer. However, then we would actually deal with a text classification task and not a seq2seq task as the notebook claims. Now I wonder if the dataset really fits the purpose.

Can you reproduce these errors? WDYT?

[BenjaminBossan]

ping @dhruvildarji

[dhruvildarji]

Thanks for testing this @BenjaminBossan! I've reproduced both issues and pushed a fix.

Root cause: With zeroshot/twitter-financial-news-sentiment, the label feature may not expose a .names attribute (plain integer feature vs ClassLabel, depending on the datasets version). When .names raises AttributeError, the text_label column is never created, so preprocess_function receives raw ints and tokenizer() fails with a TypeError.

What I changed (latest commit):

• Added a hasattr check before accessing .names across all 5 notebooks
• Falls back to ["Bearish", "Bullish", "Neutral"] if .names is unavailable
• This ensures text_label is always a list of strings before tokenization

The seq2seq framing still holds — we're fine-tuning a T5 model to generate sentiment labels as text tokens ("Bearish"/"Bullish"/"Neutral"), which is valid for a generative seq2seq model. The dataset fits the purpose.

[BenjaminBossan]

Thanks for the updates. I can confirm that the notebooks run now. The Python script still fails, but the same update should work there too.

[BenjaminBossan]

This line also needs the same adjustment as below in the notebooks, right?

[dhruvildarji]

Fixed peft_adalora_seq2seq.py as well — same hasattr fallback applied on line 32. Should be good to go now.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[BenjaminBossan]

Thanks for fixing the dataset in the examples, LGTM.