README.md

---
language:
- en
dataset_info:
  features:
  - name: instruction
    dtype: string
  - name: context
    dtype: string
  - name: response
    dtype: string
  - name: category
    dtype: string
  - name: messages
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  splits:
  - name: train
    num_bytes: 34692013
    num_examples: 15011
  download_size: 15166632
  dataset_size: 34692013
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
task_categories:
- question-answering
- text2text-generation
---
# Dataset Card for "databricks-dolly-15k-chatml"
## Dataset Summary
This dataset has been created by **Re:cast AI** to transform the existing dataset [databricks/databricks-dolly-15k](https://huggingface.co/datasets/databricks/databricks-dolly-15k) into a [chatml](https://huggingface.co/docs/transformers/main/en/chat_templating) friendly format for use in SFT tasks with pretrained models.



## Dataset Structure
```python
messages = [
  { "content": "You are an expert Q&A system that is trusted around the world. You always... etc.", "role": "system" }, 
  { "content": "(Optional) Context information is below.\n----------------\nVirgin Australia, the... etc.", "role": "user" }, 
  { "content": "Virgin Australia commenced services on 31 August 2000... etc.", "role": "assistant" } ]
]
```

## Usage
```python
from datasets import load_dataset
dataset = load_dataset("recastai/databricks-dolly-15k-chatml", split="train")
```

## Processing applied to original dataset
```python
INSTRUCTIONS = """You are an expert Q&A system that is trusted around the world. You always answer the user's query in a helpful and friendly way.
Some rules you always follow:
1. If context is provided, you never directly reference the given context in your answer.
2. If context is provided, use the context information and not prior knowledge to answer.
3. Avoid statements like 'Based on the context, ...' or 'The context information ...' or 'The answer to the user's query...' or anything along those lines.
4. If no context is provided use your internal knowledge to answer."""

# databricks-dolly-15k features:
# - instruction: The user query/question
# - context: (optional) context to use to help the assistant
# - response: The assistant's response to the query/question
#
key_mapping = dict(
    query = "instruction",
    context = "context",
    response = "response"
)

def process_chatml_fn(example, validation=False):
    """
    Processing specific to databricks-dolly-15k into a chat format.
    """
    user_content = (
        "(Optional) Context information is below.\n"
        "----------------\n"
        "{context}\n"
        "----------------\n"
        "Answer the following query.\n"
        "{query}\n"
    )
    assistant_content = "{response}"
    
    message = [
        {"role": "system", "content": INSTRUCTIONS},
        {"role": "user", "content": user_content.format(context=example[key_mapping['context']], query=example[key_mapping['query']])},
        {"role": "assistant", "content": assistant_content.format(response=example[key_mapping['response']])} 
    ]

    return message
```