Skip to content

README.md

Latest commit

 

History

History
43 lines (28 loc) · 2.08 KB

README.md

File metadata and controls

43 lines (28 loc) · 2.08 KB

RAISE-25 RESEARCH ARTIFACTS

  • data/ - Contains research_data.json, which is the master dataset containing all the raw numbers used in the visualizations.
  • figures/ - Contains the high-resolution PNG images generated by the scripts. These are the exact graphs used in the paper.
  • scripts/ - Contains the Python source code used to generate each figure. Each script is standalone and reproducible.

VISUALIZATION DETAILS

  1. Section 3: Attention Mechanism (section_3_attention.py) Visualizes the internal self-attention weights of a Transformer model. It demonstrates how the model resolves pronoun ambiguity (e.g., linking "it" to "animal") using an attention heatmap.

  2. Section 5: Empirical Scaling Laws (section_5_scaling.py) Plots the relationship between Compute (FLOPs) and Loss. It demonstrates the "Power Law," proving that performance improvements are predictable functions of increased compute resources.

  3. Section 5B: Emergent Capabilities (section_5b_emergence.py) Contrasts the smooth scaling laws with "Phase Transitions." It shows how specific capabilities (like arithmetic) remain near-zero until a critical model size is reached, appearing suddenly.

  4. Section 6: Forecasting (section_6_forecast.py) Projects future benchmark performance (MMLU) using historical trend lines. It establishes a "Cone of Uncertainty" to model best-case vs. stagnation scenarios for policy analysis.

  5. Section 7: The Trust Paradox (section_7_trust.py) A dual-axis chart illustrating the inverse relationship between Model Accuracy and Human Verification. It quantifies the risk of over-reliance as models become more capable.

  6. Section 8: Economic Impact (section_8_economic.py) Visualizes the "Skill-Leveling" effect. The data shows that LLM tools provide a significantly larger performance boost to low-skill workers compared to high-skill workers, reducing the performance gap.

HOW TO RUN

  1. Install the required Python libraries:

pip install matplotlib seaborn pandas numpy

  1. Run the master script to generate all figures:

python run_all.py

REQUIREMENTS

  • Python 3.8+
  • matplotlib
  • seaborn
  • pandas
  • numpy