Datasets:
Update README.md
Browse filesAdded dataset card.
README.md
CHANGED
|
@@ -1,3 +1,53 @@
|
|
| 1 |
---
|
| 2 |
license: cc-by-4.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: cc-by-4.0
|
| 3 |
+
task_categories:
|
| 4 |
+
- other
|
| 5 |
+
tags:
|
| 6 |
+
- agents
|
| 7 |
+
- evaluation
|
| 8 |
+
- benchmarks
|
| 9 |
+
- leaderboards
|
| 10 |
+
size_categories:
|
| 11 |
+
- 10K<n<100K
|
| 12 |
---
|
| 13 |
+
|
| 14 |
+
# HAL Agent Evaluations (Cleaned)
|
| 15 |
+
|
| 16 |
+
This repository contains the dataset used in the paper [Agent Evaluation Reliability: More Tasks Won't (Always) Fix an Agent Leaderboard](https://arxiv.org/abs/2610.00651). It is a cleaned, preprocessed version of the [HAL Leaderboard traces](https://huggingface.co/datasets/agent-evals/hal_traces), released as a single flat CSV.
|
| 17 |
+
|
| 18 |
+
The data covers **+32k task-level evaluations** across **9 benchmarks**. Each benchmark is run with 2–3 agent scaffolds and 13–34 models.
|
| 19 |
+
|
| 20 |
+
## Dataset Summary
|
| 21 |
+
|
| 22 |
+
Counts are unique levels of each facet.
|
| 23 |
+
|
| 24 |
+
| Benchmark | Models | Tasks | Agent Scaffolds | Model–Scaffold Pairs |
|
| 25 |
+
|-------------------------|:------:|:-----:|:---------------:|:--------------------:|
|
| 26 |
+
| AssistantBench | 18 | 33 | 2 | 30 |
|
| 27 |
+
| CORE-Bench Hard | 34 | 45 | 3 | 56 |
|
| 28 |
+
| GAIA | 21 | 165 | 2 | 34 |
|
| 29 |
+
| Online-Mind2Web | 13 | 300 | 2 | 23 |
|
| 30 |
+
| SciCode | 17 | 65 | 3 | 37 |
|
| 31 |
+
| ScienceAgentBench | 19 | 102 | 2 | 25 |
|
| 32 |
+
| SWE-bench Verified Mini | 24 | 50 | 2 | 26 |
|
| 33 |
+
| τ-bench Airline | 23 | 50 | 3 | 42 |
|
| 34 |
+
| USACO | 13 | 307 | 2 | 14 |
|
| 35 |
+
|
| 36 |
+
## Data Fields
|
| 37 |
+
|
| 38 |
+
| Column | Type | Description |
|
| 39 |
+
|-------------|--------|----------------------------------------------|
|
| 40 |
+
| `benchmark` | string | Benchmark name |
|
| 41 |
+
| `model` | string | Underlying LLM |
|
| 42 |
+
| `scaffold` | string | Agent scaffold / harness |
|
| 43 |
+
| `task_id` | string | Task identifier within the benchmark |
|
| 44 |
+
| `score` | float | Task outcome (e.g., 0/1 success) |
|
| 45 |
+
| ... | ... | ... |
|
| 46 |
+
|
| 47 |
+
## Usage
|
| 48 |
+
|
| 49 |
+
```python
|
| 50 |
+
from datasets import load_dataset
|
| 51 |
+
|
| 52 |
+
ds = load_dataset("razam2/hal-response-matrix")
|
| 53 |
+
```
|