datasets

Guide for dataset construction — SupervisedDatasetBuilder, RLDatasetBuilder, ChatDatasetBuilder, and custom dataset creation from JSONL, HuggingFace, or conversation data. Use when the user asks about

SKILL.md

GitHub

Datasets

The cookbook uses the builder pattern for datasets: a *DatasetBuilder (config) builds a *Dataset (runtime).

Reference

Read these for details:

  • tinker_cookbook/supervised/types.py — SupervisedDatasetBuilder, ChatDatasetBuilder, ChatDatasetBuilderCommonConfig
  • tinker_cookbook/supervised/data.py — Dataset construction helpers, FromConversationFileBuilder
  • tinker_cookbook/rl/types.py — RLDatasetBuilder, RLDataset
  • docs/training-sampling.mdx — Data preparation basics

Supervised datasets

ChatDatasetBuilderCommonConfig

Shared config for all chat-based dataset builders:

from tinker_cookbook.supervised.types import ChatDatasetBuilderCommonConfig
from tinker_cookbook.renderers import TrainOnWhat

common_config = ChatDatasetBuilderCommonConfig(
    model_name_for_tokenizer="meta-llama/Llama-3.1-8B",
    renderer_name="llama3",
    max_length=32768,        # Max sequence length
    batch_size=128,          # Tokens per batch
    train_on_what=TrainOnWhat.ALL_ASSISTANT_MESSAGES,
)

Built-in datasets

from tinker_cookbook.recipes.chat_sl.chat_datasets import NoRobotsBuilder, Tulu3Builder

dataset = NoRobotsBuilder(common_config=common_config)
dataset = Tulu3Builder(common_config=common_config)

Custom JSONL file

from tinker_cookbook.supervised.data import FromConversationFileBuilder

dataset = FromConversationFileBuilder(
    common_config=common_config,
    file_path="/path/to/data.jsonl",
    test_size=100,       # Hold out 100 examples for eval
    shuffle_seed=42,
)

JSONL format — each line is a conversation:

{"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}

See tinker_cookbook/example_data/conversations.jsonl for the expected format.

From HuggingFace datasets

from tinker_cookbook.supervised.data import SupervisedDatasetFromHFDataset

dataset = SupervisedDatasetFromHFDataset(
    hf_dataset=hf_dataset,
    batch_size=128,
    map_fn=lambda example: conversation_to_datum(
        example["messages"], renderer, max_length, train_on_what
    ),
)

Low-level datum construction

from tinker_cookbook.supervised.data import conversation_to_datum

# Full pipeline: messages → datum
datum = conversation_to_datum(messages, renderer, max_length, train_on_what)

# Or step by step:
model_input, weights = renderer.build_supervised_example(messages)
datum = datum_from_model_input_weights(model_input, weights, max_length)

RL datasets

RL datasets return batches of EnvGroupBuilder objects. See the /environments skill for details.

@chz.chz
class MyRLDatasetBuilder(RLDatasetBuilder):
    batch_size: int = 128
    group_size: int = 4

    async def __call__(self) -> tuple[RLDataset, RLDataset | None]:
        # Return (train_dataset, optional_test_dataset)
        ...

DPO datasets

DPO uses comparison pairs (chosen vs rejected):

from tinker_cookbook.preference.dpo_datasets import DPODatasetBuilderFromComparisons

dataset = DPODatasetBuilderFromComparisons(
    common_config=common_config,
    comparison_builder=HHHComparisonBuilder(),
)

See tinker_cookbook/preference/dpo_datasets.py and tinker_cookbook/recipes/preference/datasets.py.

Common pitfalls

  • Always use ChatDatasetBuilderCommonConfig for consistent tokenizer/renderer setup
  • batch_size is in tokens, not examples — larger sequences mean fewer examples per batch
  • Custom JSONL must match the format in example_data/conversations.jsonl
  • Use test_size to hold out evaluation data from the same distribution
  • Dataset builders must be serializable (@chz.chz) for config persistence and sweeps