File size: 4,776 Bytes
65ac36f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
---
license: cc-by-4.0
language:
- en
pretty_name: GAIA-Agent Practice Set (16 tool-use tasks)
size_categories:
- n<1K
task_categories:
- question-answering
tags:
- agents
- tool-use
- evaluation
- gaia
- smolagents
configs:
- config_name: default
  data_files:
  - split: test
    path: data/practice.jsonl
---

# GAIA-Agent Practice Set

**16 original, GAIA-Level-1-style tasks with verifiable answers, for regression-testing tool-using agents.** Each task can only be solved by *using* something: a spreadsheet, a CSV, a JSON export, a PDF, a Python file, an audio recording, an image, or a multi-hop Wikipedia lookup. Answers are short and graded by exact or quasi-exact match, as in [GAIA](https://huggingface.co/papers/2311.12983).

It was built for [**GAIA Agent**](https://github.com/bakulbadwal/gaia-agent), my agent for the final assignment of the [Hugging Face Agents Course](https://huggingface.co/learn/agents-course) (Unit 4). That course's scoring API returns one aggregate number and never says which answers were wrong. A local set with known answers is what lets you debug, measure reliability (pass^k) and run ablations without touching the benchmark.

> **This is not GAIA.** None of these tasks are taken or paraphrased from GAIA or from the course's 20 questions. Every task was written for this set.

## What's inside

| Skill area | Tasks | Input files |
|---|---|---|
| Spreadsheets (filter, multi-sheet group-by, dates) | 2 | `.xlsx` |
| CSV (time windows, missing values, ranking) | 2 | `.csv` |
| Code (run it, or reason about it backwards) | 2 | `.py` |
| Audio transcription + arithmetic | 2 | `.mp3` |
| Vision (chess diagram, chart reading) | 2 | `.png` |
| Documents and nested data | 2 | `.pdf`, `.json` |
| Multi-hop Wikipedia lookups | 4 | none |

Answer formats mix numbers, single words and ordered comma lists, so the format rules are tested too.

### Fields

| Field | Meaning |
|---|---|
| `task_id` | e.g. `p01-xlsx-late-loans` (`p` = file task, `w` = Wikipedia task) |
| `question` | the task, as given to the agent |
| `file_name` | path to the attached file under `files/`, or null |
| `final_answer` | gold answer |
| `answer_format` | `number`, `string` or `ordered_list` |
| `skills` | tags such as `file:xlsx`, `date-arithmetic`, `multi-hop` |
| `source_type` | `generated` or `wikipedia` |
| `source` | JSON: for generated tasks, how the gold is derived from the file; for Wikipedia tasks, each hop's URL, revision id, retrieval date (2026-10-10) and supporting sentence |
| `canary` | a canary string (below) |

## How the gold answers are known

- **Generated tasks (12):** `generator/build_practice.py` builds every file deterministically, and each gold answer is recomputed from the file itself. The project's test suite re-derives every gold answer from the generated files.
- **Wikipedia tasks (4):** each hop records the page URL, the exact revision id and the supporting sentence. Answers are pinned to those revisions; a later edit to Wikipedia could change the live page.

## Reference results

GAIA Agent (smolagents CodeAgent, free open-weight NVIDIA Nemotron 3 models via OpenRouter, 2026-10-10):

| Config | Accuracy | Notes |
|---|---|---|
| Baseline, 4 identical runs | 95.3% mean · **pass@4 100% · pass^4 93.8%** | only the chess diagram is flaky |
| Without audio and image tools | 68.8% | |
| Without web tools | 93.8% | the 4 Wikipedia answers are within the model's own knowledge, so this set **does not test retrieval** |
| Manager + 2 specialists | 93.8% | +47% tokens |

With 16 tasks one task is 6.25 points, so treat differences as directional. Full details are in the [GAIA Agent repo](https://github.com/bakulbadwal/gaia-agent).

## Limitations

- **Small:** 16 tasks is a regression suite, not a benchmark.
- **Memorizable web tasks:** strong models answer the Wikipedia questions without searching (see the table above).
- **Synthetic files:** the inputs are clean and generated, so they are easier than real-world documents.

## Licensing and attribution

- The task text, gold answers and generated files (xlsx, csv, json, pdf, png, py) are released under **CC BY 4.0**.
- The supporting sentences in `source` quote English Wikipedia (CC BY-SA 4.0), each with its page URL and revision id.
- The two `.mp3` files were synthesized with macOS system text-to-speech voices. Check Apple's terms before any commercial use of the audio itself.

## Canary

Please don't train on this set. Its canary string, for filtering it out of training corpora, is:

`GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec`

## Citation

```
Badwal, B. (2026). GAIA-Agent Practice Set: 16 tool-use tasks for regression-testing agents.
https://huggingface.co/datasets/bacool/gaia-agent-practice-set
```