essobi/trl_gair_lima

GAIR/LIMA TRL Dataset

This dataset is a transformed version of the GAIR/LIMA dataset, specifically formatted for training with TRL (Transformers Reinforcement Learning).

Dataset Summary

This dataset contains conversations from the GAIR/LIMA dataset, transformed into a message-based format suitable for conversational AI training and reinforcement learning.

Key Transformations

• Converted from Q&A format to multi-turn conversation format
• Applied message-level transformations (HTML tag removal, whitespace normalization)
• Filtered conversations based on quality criteria (length checks, content validation)
• Preserved original metadata for reference and traceability

Dataset Structure

Each example contains:

• messages: List of conversation turns, each with:
  • role: Either "user" or "assistant"
  • content: Message text
• metadata: Original and derived metadata including:
  • original_id: ID from original LIMA dataset
  • answers_count: Number of answers in original format
  • original_data: Preserved original fields

Splits

• train: Training split (filtered LIMA train set)
• test: Test split (filtered LIMA test set)

Creation Process

This dataset was created using the transformation pipeline defined in the ProblemGenerationAgent repository:

1. Loading: GAIR/LIMA dataset loaded from HuggingFace Hub
2. Transformation: Message-level transformations applied (HTML removal, whitespace normalization)
3. Filtering: Conversations filtered based on quality criteria:
   • Minimum conversation length: 2 messages
   • Minimum answer length: 50 characters
   • Maximum answer length: 5000 characters
4. Output: Transformed conversations in message format

See transformation_report.json for detailed statistics about the transformation process.

See rejection_log.json for conversations that did not pass filtering criteria.

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("essobi/trl_gair_lima")

# Access examples
for example in dataset["train"]:
    messages = example["messages"]
    metadata = example["metadata"]

License

This dataset follows the same license as the original GAIR/LIMA dataset: CC-BY-NC-4.0

Citation

Original GAIR/LIMA dataset:

@article{zhou2023lima,
  title={LIMA: Less Is More for Alignment},
  author={Zhou, Chunting and Liu, Pengfei and Xu, Puxin and others},
  journal={arXiv preprint arXiv:2305.11206},
  year={2023}
}

Disclaimer

This transformed dataset is provided as-is for research and educational purposes. The original LIMA dataset is subject to its own license and terms of use.