razam2/hal-response-matrix

HAL Agent Evaluations Response Matrix

A cleaned, task-level response matrix built from the HAL Leaderboard traces, released as a single flat CSV. This is the dataset used in Agent Evaluation Reliability: More Tasks Won't (Always) Fix an Agent Leaderboard. The data covers +32k task-level evaluations across 9 benchmarks. Each benchmark is run with 2–3 agent scaffolds and 13–34 models.

Usage

Each row is one evaluation: a single model, running in a single scaffold, attempting a single task. An agent is identified by the combination of Model_Name, Scaffold_Name, and Reasoning_Effort.

Column	Type	Description
Run_id	string	Id associated with the original hal_leaderboard trace.
Benchmark	string	Benchmark name
Task_id	string	Task ID within the benchmark
Model_Name	string	Underlying LLM
Scaffold_Name	string	Agent scaffold (harness)
Reasoning_Effort	string	Reasoning-effort setting (low, medium, high); empty if N/A
Score	float	Task score in [0, 1]

Load it with the datasets library, then pivot to get a response matrix for a single benchmark, with agents as rows and tasks as columns:

from datasets import load_dataset

df = load_dataset("razam2/hal-response-matrix")["train"].to_pandas()

gaia = df[df["Benchmark"] == "GAIA"]
matrix = gaia.pivot_table(
    index=["Model_Name", "Scaffold_Name"],
    columns="Task_id",
    values="Score",
)