Why I got a list instead of `datasets.arrow_dataset.Column`？

[septemberlemon]

[John6666]

I think it’s the design of datasets library. If you explicitly want to convert, you can also use the .to_*** functions.

# deps: pip install datasets pyarrow pandas
# docs:
# - Column return on column-name indexing: https://huggingface.co/docs/datasets/en/access
# - New Column object in releases: https://github.com/huggingface/datasets/releases
# - Access underlying Arrow table: https://discuss.huggingface.co/t/datasets-arrow-help/18880
# - pyarrow.Table.column API: https://arrow.apache.org/docs/python/generated/pyarrow.Table.html

from datasets import Dataset
import pyarrow as pa

def to_conversations(batch):
    convs = []
    for p, s in zip(batch["problem"], batch["generated_solution"]):
        convs.append(
            [{"role": "user", "content": p},
             {"role": "assistant", "content": s}]
        )
    return {"conversations": convs}

# --- minimal toy data ---
base = Dataset.from_dict({
    "problem": ["P1", "P2", "P3"],
    "generated_solution": ["S1", "S2", "S3"],
})

ds = base.map(to_conversations, batched=True)

print("=== REPRO: column-name indexing returns Column ===")
col = ds["conversations"]
print("type(ds['conversations']) =", type(col))              # datasets.arrow_dataset.Column
print("col[0] =", col[0])                                    # first conversation
assert "datasets.arrow_dataset.Column" in str(type(col))     # expected on modern versions

print("\n=== FIX 1: get Python list when you need it ===")
as_list = list(ds["conversations"])                          # materialize as plain list
print("type(list(ds['conversations'])) =", type(as_list))
print("as_list[0] =", as_list[0])
assert isinstance(as_list, list)

print("\n=== FIX 2: get the Arrow column when you need it ===")
arrow_col = ds.data.column("conversations")                  # pyarrow.ChunkedArray
print("type(ds.data.column('conversations')) =", type(arrow_col))
assert isinstance(arrow_col, pa.ChunkedArray)

print("\n=== Reference: row-first vs column-first access ===")
print("row-first type:", type(ds[0]["conversations"]))       # Python object for a single row
print("column-first type:", type(ds["conversations"]))       # Column wrapper
"""
=== REPRO: column-name indexing returns Column ===
type(ds['conversations']) = <class 'datasets.arrow_dataset.Column'>
col[0] = [{'content': 'P1', 'role': 'user'}, {'content': 'S1', 'role': 'assistant'}]

=== FIX 1: get Python list when you need it ===
type(list(ds['conversations'])) = <class 'list'>
as_list[0] = [{'content': 'P1', 'role': 'user'}, {'content': 'S1', 'role': 'assistant'}]

=== FIX 2: get the Arrow column when you need it ===
type(ds.data.column('conversations')) = <class 'pyarrow.lib.ChunkedArray'>

=== Reference: row-first vs column-first access ===
row-first type: <class 'list'>
column-first type: <class 'datasets.arrow_dataset.Column'>
"""

[septemberlemon]

I don’t very understand, will it convert to Column on the earth? what determines the type of newly added columns?

[septemberlemon]

[John6666]

what determines the type of newly added columns?

Basically by added data('s type) itself or Features if specified.

---

ds["formatted_conversations"] returns a Column view. Nothing is converted; it exposes the Arrow-backed column. Hugging Face documents that column-name indexing returns a Column object you can index like a list. (Hugging Face)

Type of a newly added column is set as follows:

• You specify it. Pass a Features schema when creating or mapping. That schema becomes the column’s Arrow type. You can later change it with cast or cast_column. (Hugging Face)
• If you do not specify it, Datasets infers the type from the Python values your map returns. Inference is Arrow-based. (Hugging Face)
• Complex returns like list[dict{...}] become nested features such as Sequence({...}). Features define column names and types. (Hugging Face)
• The dataset is backed by a PyArrow Table; low-level access is via ds.data.column("col") which yields a ChunkedArray. (Hugging Face)

Minimal patterns:

# control the new column type explicitly
from datasets import Features, Sequence, Value
features = Features({
    "formatted_conversations": Sequence({"text": Value("string"), "length": Value("int32")})
})
ds = ds.map(fn, batched=False, features=features)  # schema fixed by you

# if already created, change just one column's feature
ds = ds.cast_column("formatted_conversations",
                    Sequence({"text": Value("string"), "length": Value("int32")}))  # cast if compatible

# access choices
col_view = ds["formatted_conversations"]      # Column view
arrow_arr = ds.data.column("formatted_conversations")  # pyarrow.ChunkedArray
py_list  = list(ds["formatted_conversations"])         # plain list

Sources: column access and Column view, features and schema control, casting columns, Arrow backing and column() API. (Hugging Face)

[septemberlemon]

def generate_conversation(examples):
    problems  = examples["problem"]
    solutions = examples["generated_solution"]
    conversations = []
    for problem, solution in zip(problems, solutions):
        conversations.append([
            {"role" : "user",      "content" : problem},
            {"role" : "assistant", "content" : solution},
        ])
    return { "conversations": conversations, }
print(type(reasoning_dataset.map(generate_conversation,batched=True)[conversations ]))

that’s clear, but it did not use features parameter in the code above, why it still got a list instead of Column？

[John6666]

Oh… It seems the behavior depends on the version of the datasets library…

---

It’s version behavior, not features=.

• Why you saw a list: older datasets returned a Python list for ds["col"]. Newer versions return a datasets.arrow_dataset.Column view. The features= argument never controls this accessor; it only sets schema. (Hugging Face)

• What it is now: ds["col"] → Column view backed by a PyArrow table. The dataset is Arrow-backed. (Hugging Face)

• What sets the type of new columns:

  1. Explicit schema you pass via features= in map or later via cast_column/cast.
  2. Otherwise Arrow infers from the Python values your function returns. This becomes dataset.features. (Hugging Face)

• Make results consistent regardless of version:

  import datasets, pyarrow as pa
  print(datasets.__version__)
  
  col_view = ds["conversations"]              # Column on new versions, list on old
  as_list  = list(ds["conversations"])        # always a Python list
  arrow_ca = ds.data.column("conversations")  # always a pyarrow.ChunkedArray
  

  The Arrow interop is stable because the dataset is a PyArrow table underneath. (Hugging Face)

• Tip for nested data: Returning list[dict] from map yields a nested Sequence(struct{...}) feature unless you override with features=. Check with ds.features. (Hugging Face)

If you need Column everywhere, upgrade datasets; if you need lists, wrap with list(...).

[septemberlemon]

The datasets’s version is “4.0.0”, I use it in unsloth’s official notebook, is that very old?
in myself’s environment the version is 4.2.0
so u mean if the same code was run in myself’s environment, I’ll get a Column？
***
I’ve tried, u r right

[septemberlemon]

wait, I tried it in unsloth’s notebook again and it become Column!

but I really remember that I’ve ever got a list type and the top screenshot can prove

god, maybe there is some halloween ghost who’s trick me

[John6666]

If you want to avoid ambiguity in data types, it’s probably better to explicitly cast them…
It’s too random.