Dataset

API reference for the Dataset and Conversation classes.

Dataset

from aevyra_verdict import Dataset

Dataset.from_jsonl(path, name=None, format="auto")

Load a dataset from a JSONL file.

dataset = Dataset.from_jsonl("data.jsonl")
dataset = Dataset.from_jsonl("data.jsonl", name="my-dataset")

# Explicit format
dataset = Dataset.from_jsonl("sharegpt_data.jsonl", format="sharegpt")
dataset = Dataset.from_jsonl("alpaca_data.jsonl", format="alpaca")

Parameter	Default	Description
path	—	Path to the JSONL file.
name	filename stem	Display name for the dataset.
format	"auto"	Input format: "auto", "openai", "sharegpt", or "alpaca".

Dataset.from_list(items, name="inline", format="auto")

Create a dataset from a list of dicts (same schema as JSONL lines).

dataset = Dataset.from_list([
    {"messages": [{"role": "user", "content": "Hello"}], "ideal": "Hi"},
])

# ShareGPT inline
dataset = Dataset.from_list(sharegpt_records, format="sharegpt")

Parameter	Default	Description
items	—	List of records in OpenAI, ShareGPT, or Alpaca format.
name	"inline"	Display name for the dataset.
format	"auto"	Input format: "auto", "openai", "sharegpt", or "alpaca".

### `dataset.filter(**kwargs)`

Return a new dataset containing only conversations where metadata matches all given key-value pairs.

```python
hard = dataset.filter(difficulty="hard")
hard_reasoning = dataset.filter(difficulty="hard", category="reasoning")

dataset.summary()

Return a dict with name, sample count, whether ideals are present, and metadata keys.

{
    "name": "data",
    "num_conversations": 50,
    "has_ideals": True,
    "metadata_keys": ["category", "difficulty"]
}

dataset.has_ideals()

Return True if every conversation has an ideal field.

---

Conversation

Each item in a dataset is a Conversation:

Property	Type	Description
messages	list[Message]	The conversation messages.
ideal	str | None	The reference answer.
metadata	dict	Arbitrary metadata.
prompt_messages	list[dict]	Messages as plain dicts, ready to send to a provider.
last_user_message	str | None	The last user message in the conversation.