license: cc-by-4.0
pretty_name: Gradient Decomposition Assay
language:

• en
  tags:
• llm-evaluation
• ai-safety
• prompt-evaluation
• llm-as-judge
• red-teaming
  configs:
• config_name: structured_ordered_flagged
  default: true
  data_files: "data/structured_ordered_flagged.csv"
• config_name: structured_ordered_valid_only
  data_files: "data/structured_ordered_valid_only.csv"
• config_name: raw_telemetry_ordered_flagged
  data_files: "data/raw_telemetry_ordered_flagged.csv"
• config_name: vector_summary_valid
  data_files: "data/vector_summary_valid.csv"
• config_name: model_vector_summary_valid
  data_files: "data/model_vector_summary_valid.csv"
• config_name: invalid_runs
  data_files: "data/invalid_runs.csv"

Gradient Decomposition Assay (GDA)

This repository contains the CSV corpus and summary tables for the Gradient Decomposition Assay, an exploratory behavioral evaluation of how frontier language models respond to an eight-vector prompt manifold ranging from benign technical tasks to adversarial compression and counterfactual/narrative reframing.

Why this repository uses multiple configurations

The CSV files in this dataset are not all the same table. The row-level structured telemetry, raw telemetry, vector summaries, model-by-vector summaries, and invalid-run audit have different schemas. Hugging Face's CSV dataset builder expects files inside one configuration to share the same columns, so each table is declared as a separate configuration in the YAML block at the top of this README.

Configurations

• structured_ordered_flagged: row-level structured telemetry, ordered by vector/model/iteration, with invalid-run flags retained.
• structured_ordered_valid_only: row-level structured telemetry with invalid/malformed runs removed.
• raw_telemetry_ordered_flagged: raw model outputs and evaluator JSON outputs, ordered and flagged.
• vector_summary_valid: vector-level means over valid rows.
• model_vector_summary_valid: model-by-vector means over valid rows.
• invalid_runs: audit table for malformed, parse-error, or otherwise invalid rows.

Loading examples

from datasets import load_dataset

# Default row-level table with invalid-run flags retained
ds = load_dataset("devinendorphin/Gradient-Decomposition-Assay", "structured_ordered_flagged")

# Vector-level summary table
summary = load_dataset("devinendorphin/Gradient-Decomposition-Assay", "vector_summary_valid")

# Raw outputs and evaluator payloads
raw = load_dataset("devinendorphin/Gradient-Decomposition-Assay", "raw_telemetry_ordered_flagged")

Metrics

The structured tables include evaluator-assigned scores for:

• Phi_Content
• Phi_Form
• Phi_Specificity
• Safety_Drag
• Self_Audit
• Refusal_Intensity
• Boilerplate_Intensity

These are LLM-judge ratings and should be treated as exploratory behavioral measurements, not ground-truth measurements of latent model state or internal reasoning capacity.

Data quality notes

The original design matrix contained 2,000 runs. The primary valid-only summaries exclude invalid or malformed rows; the invalid-run table is retained separately for auditability.

Citation

If using this dataset, cite the accompanying manuscript or repository description supplied with the release.