spivi87/alephbert-intent-he-samples

AlephBERT Hebrew Shopping Intent — Sample Dataset

A 102-row sample of the synthetic training data used to fine-tune spivi87/alephbert-intent-he.

Schema

Column	Type	Description
text	string	Hebrew utterance
label	string	Intent label (e.g. GROCERY_REQUEST, OTHER)
label_en	string	One-line English description of the intent
label_id	int64	Numeric label id (0–16, matches the model output)

Generation

Synthetic data, generated via GPT-4o-mini paraphrasing of Hebrew seed templates. No real WhatsApp / user messages were used. Generation script and seed templates live in the standalone repo: github.com/spivi/alephbert-intent-he.

Methodology — seed-level train/test split

This sample is drawn only from the TRAIN split of the corpus used to fine-tune spivi87/alephbert-intent-he. The train/test split happens at the seed level, before paraphrasing: for every intent, 2 seeds are held out from training, and the test set contains only paraphrases of those held-out seeds. This prevents paraphrases of the same source seed from leaking across splits — a common pitfall in LLM-generated synthetic data that inflates reported accuracy.

Limitations

• 17 intent classes, all shopping/grocery-domain
• Class balance is roughly uniform but not exact
• This is a sample of the training set, not the full corpus, and not the held-out test set

License

Apache 2.0 (same as the model and the AlephBERT base).