📄 Hh Rlhf¶

This repository provides access to two different kinds of data: 1) Human preference data about helpfulness and harmlessness from Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback. These data are meant to train preference (or reward) models for subsequent RLHF training. These data are not meant for supervised training of dialogue agents. Training dialogue agents on these data is likely… See the full description on the dataset page: https://huggingface.co/datasets/Anthropic/hh-rlhf.

Tags: arxiv:2204.05862, flags:['human-feedback'], license:mit, region:us

cards.hh_rlhf

type: TaskCard
loader: 
  type: LoadHF
  path: Anthropic/hh-rlhf
preprocess_steps: 
  - splitters.small_no_dev
  - type: Strip
    field: chosen
  - type: Replace
    field: chosen
    old: "\n\n"
    new: "\n"
  - type: Split
    field: chosen
    by: "\nAssistant:"
  - type: Get
    field: chosen
    item: -1
    to_field: output_choice
  - type: Slice
    field: chosen
    stop: -1
  - type: Join
    field: chosen
    by: "\nAssistant:"
    to_field: input
  - type: Split
    field: rejected
    by: "\nAssistant:"
  - type: Get
    field: rejected
    item: -1
    to_field: output_rejected
  - type: ListFieldValues
    fields: 
      - output_choice
      - output_rejected
    to_field: choices
  - type: ShuffleFieldValues
    field: choices
  - type: Set
    fields: 
      input_type: dialog
      output_type: response
      instruction: Respond the following dialog in an helpful and harmfull way.
  - type: IndexOf
    search_in: choices
    index_of: output_choice
    to_field: output_choice
task: tasks.evaluation.preference
templates: templates.evaluation.preference.all
[source]

Explanation about TaskCard¶

TaskCard delineates the phases in transforming the source dataset into model input, and specifies the metrics for evaluation of model output.

Attributes:
loader: specifies the source address and the loading operator that can access that source and transform it into a unitxt multistream.

preprocess_steps: list of unitxt operators to process the data source into model input.

task: specifies the fields (of the already (pre)processed instance) making the inputs, the fields making the outputs, and the metrics to be used for evaluating the model output.

templates: format strings to be applied on the input fields (specified by the task) and the output fields. The template also carries the instructions and the list of postprocessing steps, to be applied to the model output.

Explanation about LoadHF¶

Loads datasets from the HuggingFace Hub.

It supports loading with or without streaming, and it can filter datasets upon loading.
Args:
path: The path or identifier of the dataset on the HuggingFace Hub. name: An optional dataset name. data_dir: Optional directory to store downloaded data. split: Optional specification of which split to load. data_files: Optional specification of particular data files to load. revision: Optional. The revision of the dataset. Often the commit id. Use in case you want to set the dataset version. streaming: Bool indicating if streaming should be used. filtering_lambda: A lambda function for filtering the data after loading. num_proc: Optional integer to specify the number of processes to use for parallel dataset loading.

Example:
Loading glue’s mrpc dataset
load_hf = LoadHF(path='glue', name='mrpc')

Explanation about Set¶

Adds specified fields to each instance in a given stream or all streams (default) If fields exist, updates them.

Args:

fields (Dict[str, object]): The fields to add to each instance.
Use ‘/’ to access inner fields

use_deepcopy (bool) : Deep copy the input value to avoid later modifications

Examples:
# Add a ‘classes’ field with a value of a list “positive” and “negative” to all streams Set(fields={“classes”: [“positive”,”negatives”]})

# Add a ‘start’ field under the ‘span’ field with a value of 0 to all streams Set(fields={“span/start”: 0}

# Add a ‘classes’ field with a value of a list “positive” and “negative” to ‘train’ stream Set(fields={“classes”: [“positive”,”negatives”], apply_to_stream=[“train”]})

# Add a ‘classes’ field on a given list, prevent modification of original list # from changing the instance. Set(fields={“classes”: alist}), use_deepcopy=True) # if now alist is modified, still the instances remain intact.

Explanation about ListFieldValues¶

Concatenates values of multiple fields into a list, and assigns it to a new field.

Explanation about IndexOf¶

For a given instance, finds the offset of value of field ‘index_of’, within the value of field ‘search_in’.

Explanation about ShuffleFieldValues¶

Shuffles a list of values found in a field.

References: templates.evaluation.preference.all, tasks.evaluation.preference, splitters.small_no_dev