---
language:
- en
license: apache-2.0
size_categories:
- n<1K
task_categories:
- text-classification
pretty_name: TL (Test vs Learn) chatbot prompts
dataset_info:
  features:
  - name: text
    dtype: string
  - name: label
    dtype:
      class_label:
        names:
          '0': learn
          '1': test
  splits:
  - name: train
    num_bytes: 47555
    num_examples: 371
  - name: test
    num_bytes: 28313
    num_examples: 262
  download_size: 50335
  dataset_size: 75868
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
tags:
- llms
- nlp
- chatbots
- prompts
---

This dataset contains manually labeled examples used for training and testing [reddgr/tl-test-learn-prompt-classifier](https://huggingface.co/reddgr/tl-test-learn-prompt-classifier), a fine-tuning of DistilBERT that classifies chatbot prompts as either 'test' or 'learn.'

Prompts labeled as 'test' (1) are those where it can be inferred that the user is:

- Presenting a problem that requires complex reasoning or arithmetic logic to resolve.
- Intentionally 'challenging' the conversational tool with a complicated question the user might know the answer to.
- Applying prompt engineering techniques such as "chain of thought" or role play.
- Presenting a highly subjective question the user makes with the purpose of testing the tool rather than learning from it or obtaining a specific unknown information.

Prompts labeled as 'instruction' (0) are those containing straightforward questions or requests where it can be inferred the user expects to learn something or obtain valuable/practical information from the interaction.

An alternative naming convention for the labels is 'problem' (test) vs 'instruction' (learn). The earliest versions of the reddgr/tl-test-learn-prompt-classifier model used a zero-shot classification pipeline for those two specific terms: instruction (0) vs problem (1).

Important note about accuracy metrics: coding questions, involving programming language syntax, are a often category of their own and are typically difficult to categorize with this dataset. This dataset contains several manually tagged coding examples, generally tagged as 'learn' (language models often mistake regular coding requests and questions with arithmetic problems).  

This dataset and the model are part of a project aimed at identifying metrics to quantitatively measure the conversational quality of text generated by large language models (LLMs) and, by extension, any other type of text extracted from a conversational context (customer service chats, social media posts...).

Relevant Jupyter notebooks and Python scripts that use this dataset and related datasets and models can be found in the following GitHub repository:
[reddgr/chatbot-response-scoring-scbn-rqtl](https://github.com/reddgr/chatbot-response-scoring-scbn-rqtl)
## Labels:
- **LABEL_0**: Learn (instruction)
- **LABEL_1**: Test (problem)