README.md

metadata

language:
  - en
license: mit
size_categories:
  - 10K<n<100K
task_categories:
  - question-answering
  - text-analysis
tags:
  - knowledge-coupling
  - llama2
  - hotpotqa
  - multi-hop-reasoning
  - gradient-analysis
  - ripple-effects

Knowledge Coupling Analysis on HotpotQA Dataset

Dataset Description

This dataset contains the results of a comprehensive knowledge coupling analysis performed on the HotpotQA dataset using LLaMA2-7B model. The analysis investigates how different pieces of knowledge interact within the model's parameter space through gradient-based coupling measurements.

Research Overview

• Model: meta-llama/Llama-2-7b-hf (layers 28-31 focused analysis)
• Dataset: HotpotQA (train + dev splits, 97,852 total samples)
• Method: Gradient-based knowledge coupling via cosine similarity
• Target Layers: model.layers.28-31.mlp.down_proj (semantically rich layers)

Key Findings

The analysis revealed:

• Mean coupling score: 0.0222 across all knowledge piece pairs
• High coupling pairs (≥0.4 threshold): Critical for ripple effect prediction
• Layer-specific analysis focusing on MLP down-projection layers
• Comprehensive gradient analysis with 180,355,072 dimensions per knowledge piece

Files Description

Core Results

• global_analysis_results.json: Comprehensive analysis summary with statistics
• all_knowledge_pieces.json: Complete set of processed knowledge pieces (92MB)
• all_coupling_pairs.csv: All pairwise coupling measurements (245MB)

Supporting Files

• dataset_info.json: Dataset statistics and conversion details
• coupling_analysis_config.json: Analysis configuration and parameters

Usage

from datasets import load_dataset

# Load the knowledge coupling results
dataset = load_dataset("your-username/hotpotqa-knowledge-coupling")

# Access global analysis results
global_results = dataset["global_analysis"]

# Access knowledge pieces
knowledge_pieces = dataset["knowledge_pieces"]

# Access coupling pairs
coupling_pairs = dataset["coupling_pairs"]

Citation

If you use this dataset in your research, please cite:

@dataset{hotpotqa_knowledge_coupling,
  title={Knowledge Coupling Analysis on HotpotQA Dataset using LLaMA2-7B},
  author={[Your Name]},
  year={2024},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/your-username/hotpotqa-knowledge-coupling}
}

Technical Details

• Gradient Computation: ∇_θ log P(answer|question) for cloze-style questions
• Coupling Measurement: Cosine similarity between L2-normalized gradients
• Memory Optimization: Focused on layers 28-31 to handle GPU memory constraints
• Hardware: NVIDIA A40 GPU (46GB VRAM)

License

This dataset is released under the MIT License. The original HotpotQA dataset follows its respective licensing terms.