Data Mapper

Introduction

Data Mapper enables users to transform data from an output record generator into SFT or DPO format, based on configuration preferences.

Key Features

1. YAML-Based Transformation Control
   Users can select the transformation format (SFT/DPO) by modifying the required and type keys under the oasst_mapper section in the configuration file.

2. Automated Output Validation
   Once a transformation type is chosen, the output is automatically schema-validated—no manual validation required.

---

Default Behavior

By default, the Data Mapper is disabled. To enable it, add the following runtime argument:

--oasst True

The default configuration is specified in the configuration.yaml file:

oasst_mapper:
  required: "yes"
  type: "sft"
  intermediate_writing: "no"

• required: Indicates whether transformation is mandatory ("yes").
• type: Defines the transformation format ("sft" or "dpo").
• intermediate_writing: Enables saving intermediate files for debugging ("yes" or "no").

---

Usage Illustration

Given an input record like the following:

{
  "id": 602,
  "conversation": [
    {
      "role": "user",
      "content": "How would you draft a Python function that identifies the initial recurring character in a specified string?"
    },
    {
      "role": "assistant",
      "content": "def first_repeated_char(s):\n    seen = set()\n    for char in s:\n        if char in seen:\n            return char\n        seen.add(char)\n    return None"
    }
  ],
  "taxonomy": [{"category": "Coding", "subcategory": ""}],
  "annotation_type": ["mistral-large"],
  "language": "en",
  "tags": ["mbpp", "reannotate", "self-critique"]
}

To Transform

You can run the Data Mapper in two modes:

1. Default Mode

Run with:

--oasst True

This uses the configuration specified in configuration.yaml:

oasst_mapper:
  required: "yes"
  type: "sft"
  intermediate_writing: "no"

2. Overridden Mode

To override the default behavior, modify the output_config section in your graph_config.yaml file like so:

output_config:
  oasst_mapper:
    required: "yes"
    type: "sft"
    intermediate_writing: "yes"

💡 Set intermediate_writing to "yes" to store intermediate files before transformation (useful for debugging). If not provided or set to any other value, it defaults to "no".

Output Example

The transformed output will follow the selected format (e.g., SFT), like this:

[
  {
    "conversation_id": "conv_8513aa73",
    "message_id": "msg_1_59a96096",
    "parent_id": null,
    "root_message_id": "msg_1_59a96096",
    "message_level": 1,
    "role": "user",
    "content": "...",
    ...
  },
  {
    "conversation_id": "conv_8513aa73",
    "message_id": "msg_2_e8b221e8",
    "parent_id": "msg_1_59a96096",
    "root_message_id": "msg_1_59a96096",
    "message_level": 2,
    "role": "assistant",
    "content": "...",
    ...
  }
]

---

Rules for Using Data Mapping

1. To run in default mode, simply use the runtime flag --oasst True.
2. If you want to override defaults, define oasst_mapper inside output_config in the config file.
3. If oasst_mapper is defined, its keys must be:
4. required: must be "yes" or "no"
5. type: must be "sft" or "dpo"
6. intermediate_writing: optional, defaults to "no" if not "yes"
7. If keys are missing or values are invalid, exceptions will be raised.

---

Additional Features

Support for custom transformation types and custom validation schemas is planned in a future release.