fix: convert Dataset Column to list before SentenceTransformer.encode

[octo-patch]

Fixes #79

Problem

SentenceTransformer.encode() sorts input sentences by length internally using numpy.int64 indices. HuggingFace Dataset column objects (e.g. data["train"]["text"]) do not support indexing with numpy.int64, which raises a TypeError during encode.

Solution

Wrap the column with list() before passing it to model.encode(), converting it to a plain Python list that supports standard numpy indexing:

# Before
train_embeddings = model.encode(data["train"]["text"], show_progress_bar=True)
test_embeddings  = model.encode(data["test"]["text"],  show_progress_bar=True)

# After
train_embeddings = model.encode(list(data["train"]["text"]), show_progress_bar=True)
test_embeddings  = model.encode(list(data["test"]["text"]),  show_progress_bar=True)

Testing

The fix matches the workaround confirmed in the issue thread. The list() conversion has zero impact on encode results — it only changes the container type before sort-by-length occurs.

[seanv507]

IMO this problem is fixed by pinning sentence transformers as mentioned in this issue
#107 (comment)

(no issue when running colab, pinning transformers, sentence transformers and peft to versions in requirements.txt)

[octo-patch]

Thanks — pinning works as a workaround, but this 2-line defensive change keeps the notebook running on whatever sentence-transformers / datasets versions a reader happens to install (e.g. fresh Colab runtimes with newer pins). It costs almost nothing and avoids users hitting the same TypeError again the next time the upstream pins drift. Happy to defer if you prefer to handle this purely via requirements.txt.