Open Platypus: A Curated Collection for LLM Logical Reasoning

A dataset from garage-bAInd, last updated on 2024-01-24, focused on improving large language model logical reasoning skills. It was used to train the Platypus2 models and is comprised of several filtered datasets including PRM800K, MATH, ScienceQA, SciBench, ReClor, and TheoremQA.

Use Cases

• Fine-tuning LLMs for mathematical problem-solving based on the MATH and TheoremQA components.
• Training models on scientific question-answering based on the ScienceQA and SciBench datasets.
• Improving logical reasoning and reading comprehension for LLMs based on the ReClor dataset.
• Developing models for chain-of-thought or process-based reasoning based on the PRM800K dataset.

Strengths

• Comprised of multiple established datasets (PRM800K, MATH, ScienceQA, SciBench, ReClor, TheoremQA) with known licenses.
• Filtered using Sentence Transformers to remove questions with a similarity above 80%, likely reducing redundancy.
• Explicitly designed and used for training the Platypus2 models.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and total size are unknown, which may limit suitability assessment.
• Data may reflect biases inherent to the source datasets, which span academic and logical reasoning contexts.

Provenance

Source

garage-bAInd on Hugging Face.

Collection Method

Curated from multiple datasets, filtered using keyword search and Sentence Transformers.

Time Range

null

Freshness

Last updated 2024-01-24 19:09:41; freshness should be verified.

Geography

null