GAIA: a benchmark for General AI Assistants
Paper • 2311.12983 • Published • 252
Datasets:
task_id stringlengths 16 24 | question stringlengths 103 300 | file_name stringlengths 18 30 ⌀ | final_answer stringlengths 1 40 | answer_format stringclasses 3
values | skills listlengths 3 5 | source_type stringclasses 2
values | source stringlengths 138 654 | canary stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|
p01-xlsx-late-loans | The attached spreadsheet is a library's loan log. Consider only loans by members in the Student tier that were checked out during March 2026. How many of those loans were returned more than 21 days after the checkout date? Loans with no return date do not count. | files/library_loans.xlsx | 8 | number | [
"file:xlsx",
"filter",
"date-arithmetic",
"count",
"strict-bound"
] | generated | {"type": "generated", "generator": "evals/practice/build_practice.py", "derivation": "Tier == Student, checkout in 2026-03, Returned not blank, (Returned - Checked Out).days > 21"} | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
p02-xlsx-multisheet-sku | The attached workbook has one sheet per warehouse. Summed across all warehouses, which SKU has the largest total quantity on hand? Answer with the SKU code only. | files/warehouse_stock.xlsx | KX-2210 | string | [
"file:xlsx",
"multi-sheet",
"group-by",
"argmax"
] | generated | {"type": "generated", "generator": "evals/practice/build_practice.py", "derivation": "concat all sheets, group by SKU, sum Qty On Hand, argmax (unique)"} | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
p03-csv-evening-trips | The attached CSV lists one day's shuttle trips. How many trips departed at or after 18:00 and took longer than 90 minutes from departure to arrival? | files/shuttle_trips.csv | 4 | number | [
"file:csv",
"datetime",
"filter",
"count",
"strict-bound"
] | generated | {"type": "generated", "generator": "evals/practice/build_practice.py", "derivation": "departure.hour >= 18 and (arrival - departure) > 90 min (overnight arrivals included)"} | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
p04-csv-pm25-ranking | The attached file has hourly PM2.5 readings from several monitoring stations; blank cells are missed readings and should be ignored, not treated as zero. Which three stations have the highest mean PM2.5? List the station names, highest mean first, separated by commas. | files/air_quality_hourly.csv | Harbour Road, Station Square, Kingsway | ordered_list | [
"file:csv",
"missing-values",
"group-by",
"ranking",
"ordered-list"
] | generated | {"type": "generated", "generator": "evals/practice/build_practice.py", "derivation": "groupby(station).mean() with NaN skipped, top 3 descending"} | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
p05-py-queue-sim | The attached Python script simulates a ticket counter and prints how many customers are waiting after 60 minutes. If the same simulation were run for 240 minutes with three counters open instead of the default (everything else unchanged), how many customers would still be waiting at the end? | files/queue_sim.py | 4 | number | [
"file:py",
"code-execution",
"modify-parameters"
] | generated | {"type": "generated", "generator": "evals/practice/build_practice.py", "derivation": "simulate(240, counters=3) from the attached module"} | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
p06-py-invert-cipher | The attached Python module defines a function encode(). Which lowercase English word w satisfies encode(w) == 'kkbeqsz'? | files/label_cipher.py | lantern | string | [
"file:py",
"code-reasoning",
"inversion"
] | generated | {"type": "generated", "generator": "evals/practice/build_practice.py", "derivation": "invert encode(): reverse, then subtract key shift and position index mod 26", "ciphertext": "kkbeqsz"} | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
p07-audio-platform | Listen to the attached station announcement. From which platform should a passenger travelling to Preston now depart? Answer with the platform number only. | files/station_announcement.mp3 | 11 | number | [
"file:audio",
"transcription",
"distractors"
] | generated | {"type": "generated", "generator": "evals/practice/build_practice.py", "derivation": "TTS script (source.tts_text): the 19:05 to Preston is cancelled; take the 19:20 from platform eleven", "tts_text": "Good evening. This is a platform alteration. The eighteen fifteen service to Carlisle will now depart from platform ni... | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
p08-audio-mean-reading | The attached recording is a technician reading out temperature measurements in degrees Celsius. What is the mean of the readings, rounded to one decimal place? | files/sensor_readings.mp3 | 20.6 | number | [
"file:audio",
"transcription",
"arithmetic",
"rounding"
] | generated | {"type": "generated", "generator": "evals/practice/build_practice.py", "derivation": "mean of the five readings in the TTS script, ROUND_HALF_UP to 0.1", "tts_text": "Sensor log for bay three, recorded at the start of the shift. All readings are in degrees Celsius. Reading one: 21.4. Reading two: 19.8. Reading three: 2... | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
p09-image-chess-mate | The attached image shows a chess position with White to move (White pieces are drawn in white, Black pieces in black; White plays up the board from rank 1). White has exactly one move that delivers checkmate immediately. Give that move in standard algebraic notation (SAN), including the mate symbol. | files/chess_diagram.png | Qe8# | string | [
"file:image",
"vision",
"chess",
"verification"
] | generated | {"type": "generated", "generator": "evals/practice/build_practice.py", "derivation": "python-chess: legal moves from source.fen whose result is checkmate (exactly one)", "fen": "6k1/1n1Q1ppp/6q1/R7/7N/B5b1/5PPP/6K1 w - - 0 1"} | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
p10-image-chart-mean | The attached bar chart shows a weather station's monthly rainfall for one year. Which month's rainfall was closest to that year's monthly average? Give the full month name. | files/rainfall_chart.png | September | string | [
"file:image",
"vision",
"chart-reading",
"arithmetic"
] | generated | {"type": "generated", "generator": "evals/practice/build_practice.py", "derivation": "argmin |value - mean(values)| over the 12 plotted values (source.values)", "values": {"January": 88, "February": 71, "March": 57, "April": 55, "May": 49, "June": 38, "July": 29, "August": 41, "September": 66, "October": 92, "November"... | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
p11-pdf-action-items | How many action items in the attached meeting minutes are owned by Priya Raman and are due before 15 November 2026? | files/procurement_minutes.pdf | 3 | number | [
"file:pdf",
"filter",
"date-comparison",
"near-miss-names",
"count"
] | generated | {"type": "generated", "generator": "evals/practice/build_practice.py", "derivation": "parse 'AIxx | ... | Owner: X | Due: D' lines; Owner == 'Priya Raman' and D < 2026-11-15"} | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
p12-json-top-customer | The attached JSON is an order export. Counting only orders whose status is "shipped", which customer spent the most in total (sum of quantity times unit price over their items)? Give the customer's full name as First Last. | files/orders_export.json | Theo Nakamura | string | [
"file:json",
"nested-data",
"filter",
"group-by",
"argmax"
] | generated | {"type": "generated", "generator": "evals/practice/build_practice.py", "derivation": "status == shipped; per customer sum(qty * unit_price); argmax (unique)"} | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
w01-wiki-element-city | A chemical element is named after the Latin name of the capital city of Sweden. What is that element's atomic number? | null | 67 | number | [
"web",
"wikipedia",
"multi-hop",
"number"
] | wikipedia | {"type": "wikipedia", "retrieved": "2026-10-10", "hops": [{"url": "https://en.wikipedia.org/wiki/Holmium", "revid": 1377857933, "retrieved": "2026-10-10", "evidence": "The element's name comes from Holmia, the Latin name for the city of Stockholm."}, {"url": "https://en.wikipedia.org/wiki/Holmium", "revid": 1377857933,... | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
w02-wiki-architect-prize | In what year was the architect who designed the Sydney Opera House awarded the Pritzker Architecture Prize? | null | 2003 | number | [
"web",
"wikipedia",
"multi-hop",
"year"
] | wikipedia | {"type": "wikipedia", "retrieved": "2026-10-10", "hops": [{"url": "https://en.wikipedia.org/wiki/Sydney_Opera_House", "revid": 1378768237, "retrieved": "2026-10-10", "evidence": "Designed by Danish architect Jørn Utzon and completed by an Australian architectural team headed by Peter Hall, the building was formally ope... | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
w03-wiki-bolivia-borders | Which countries share a land border with Bolivia? List them in alphabetical order, separated by commas. | null | Argentina, Brazil, Chile, Paraguay, Peru | ordered_list | [
"web",
"wikipedia",
"list",
"ordered-list",
"alphabetical"
] | wikipedia | {"type": "wikipedia", "retrieved": "2026-10-10", "hops": [{"url": "https://en.wikipedia.org/wiki/Bolivia", "revid": 1378338095, "retrieved": "2026-10-10", "evidence": "Bolivia is bordered by Brazil to the north and east, Paraguay to the southeast, Argentina to the south, Chile to the southwest, and Peru to the west."}]... | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
w04-wiki-mozart-river | Which river flows through the city where Wolfgang Amadeus Mozart was born? Give only the river's name, without the word "River". | null | Salzach | string | [
"web",
"wikipedia",
"multi-hop",
"entity"
] | wikipedia | {"type": "wikipedia", "retrieved": "2026-10-10", "hops": [{"url": "https://en.wikipedia.org/wiki/Wolfgang_Amadeus_Mozart", "revid": 1379458284, "retrieved": "2026-10-10", "evidence": "Born in Salzburg, Mozart quickly emerged as a child prodigy under the training of his father Leopold, a skilled pedagogue."}, {"url": "h... | GAIA-AGENT-PRACTICE CANARY 65a201c3-653f-4400-958d-b59accba6eec |
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.
It was built for GAIA Agent, my agent for the final assignment of the Hugging Face 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.pybuilds 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.
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
sourcequote English Wikipedia (CC BY-SA 4.0), each with its page URL and revision id. - The two
.mp3files 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
- Downloads last month
- 118