Registry

All Harbor datasets available as Inspect tasks. Use the search box to filter by name or description, the category chips to filter by topic, and the column headers to sort. Click a dataset’s name to open its page here, which includes a link to the dataset’s page on the Harbor registry.

Usage

CLI:

inspect eval inspect_harbor/aime_1_0 --model openai/gpt-5

Python:

from inspect_ai import eval
from inspect_harbor import aime_1_0

eval(aime_1_0(), model="openai/gpt-5")

Available Datasets

Harbor Dataset Inspect Task Description Samples
ade-bench@1.0 ade_bench_1_0 Analytics Data Engineer Bench: tasks evaluating AI agents on dbt/SQL data analytics engineering bugs. 48
aider-polyglot@1.0 aider_polyglot_1_0 A polyglot coding benchmark that evaluates AI agents’ ability to perform code editing and generation tasks across multiple programming languages. 225
aime@1.0 aime_1_0 American Invitational Mathematics Examination (AIME) benchmark for evaluating mathematical reasoning and problem-solving capabilities. Contains 60 competition-level mathematics problems from AIME 2024, 2025-I, and 2025-II competitions. 60
algotune@1.0 algotune_1_0 AlgoTune: 154 algorithm optimization tasks focusing on speedup-based scoring from the AlgoTune benchmark. 154
arc_agi_2@1.0 arc_agi_2_1_0 ARC-AGI-2: A benchmark measuring abstract reasoning through visual grid puzzles requiring rule inference and generalization. 167
autocodebench@lite200 autocodebench_lite200 Adapter for AutoCodeBench. 200
bfcl@1.0 bfcl_1_0 Berkeley Function-Calling Leaderboard: 3,641 function calling tasks for evaluating LLM tool use capabilities across simple, multiple, parallel, and irrelevance categories. 3641
bfcl_parity@1.0 bfcl_parity_1_0 BFCL parity subset: 123 stratified sampled tasks for validating Harbor adapter equivalence with original BFCL benchmark. 123
bigcodebench-hard-complete@1.0.0 bigcodebench_hard_complete_1_0_0 BigCodeBench-Hard complete benchmark adapter for Harbor - challenging Python programming tasks with reward-based verification. 145
binary-audit@1.0 binary_audit_1_0 An open-source benchmark for evaluating AI agents’ ability to find backdoors hidden in compiled binaries. 46
bird-bench@parity bird_bench_parity BIRD SQL parity subset (150 tasks, seed 42). 150
bixbench-cli@1.5 bixbench_cli_1_5 bixbench-cli - A benchmark for evaluating AI agents on bioinformatics and computational biology tasks. (Adapted for CLI execution). 205
bixbench@1.5 bixbench_1_5 BixBench - A benchmark for evaluating AI agents on bioinformatics and computational biology tasks. 205
code-contests@1.0 code_contests_1_0 A competitive programming benchmark from DeepMind that evaluates AI agents’ ability to solve algorithmic problems, covering algorithms, data structures, and competitive programming challenges. 9644
codepde@1.0 codepde_1_0 CodePDE evaluates code generation capabilities on scientific computing tasks, specifically focusing on Partial Differential Equation (PDE) solving. 5
compilebench@1.0 compilebench_1_0 Version 1.0 of CompileBench, a benchmark on real open-source projects against dependency hell, legacy toolchains, and complex build systems. 15
cooperbench@1.0 cooperbench_1_0 CooperBench: multi-agent cooperation benchmark. 652 feature pairs across 12 repos requiring two agents to coordinate via messaging. 652
crustbench@1.0 crustbench_1_0 CRUST-bench: 100 C-to-safe-Rust transpilation tasks from real-world C repositories. 100
dabstep@1.0 dabstep_1_0 DABstep: Data Agent Benchmark for Multi-step Reasoning. 450 tasks where agents analyze payment transaction data with Python/pandas to answer business questions. 450
deveval@1.0 deveval_1_0 DevEval benchmark: comprehensive evaluation of LLMs across software development lifecycle (implementation, unit testing, acceptance testing) for 21 real-world repositories across Python, C++, Java, and JavaScript. 63
ds-1000@head ds_1000_head DS-1000 is a code generation benchmark with 1000 realistic data science problems across seven popular Python libraries. 1000
evoeval@1.0 evoeval_1_0 EvoEval_difficult: 100 challenging Python programming tasks evolved from HumanEval. 100
featurebench-lite-modal@1.0 featurebench_lite_modal_1_0 FeatureBench lite split for Modal: 30 feature-implementation tasks with gpus=1 for GPU tasks (7/30). Use with -e modal. 30
featurebench-lite@1.0 featurebench_lite_1_0 FeatureBench lite split: 30 feature-implementation tasks (26 lv1 + 4 lv2) across Python repos. 30
featurebench-modal@1.0 featurebench_modal_1_0 FeatureBench full split for Modal: 200 feature-implementation tasks with gpus=1 for GPU tasks (44/200). Use with -e modal. 200
featurebench@1.0 featurebench_1_0 FeatureBench full split: 200 feature-implementation tasks across 24 Python repos. 7 tasks require Ampere+ GPU. 200
financeagent@public financeagent_public Finance Agent is a tool for financial research and analysis that leverages large language models and specialized financial tools to answer complex queries about companies, financial statements, and SEC filings. 50
gaia@1.0 gaia_1_0 GAIA (General AI Assistants): 165 validation tasks for multi-step reasoning, tool use, and multimodal question answering. 165
gpqa-diamond@1.0 gpqa_diamond_1_0 GPQA Diamond subset: 198 graduate-level multiple-choice questions in biology, physics, and chemistry for evaluating scientific reasoning. 198
gso@1.0 gso_1_0 GSO: 102 software optimization tasks focusing on performance improvement. 102
hello-world@1.0 hello_world_1_0 A simple example task to create a hello.txt file with ‘Hello, world!’ as content. 1
humanevalfix@1.0 humanevalfix_1_0 HumanEvalFix: 164 Python code repair tasks from HumanEvalPack. 164
ineqmath@1.0 ineqmath_1_0 This adapter brings IneqMath, the dev set of the first inequality-proof Q&A benchmark for LLMs, into Harbor, enabling standardized evaluation of models on mathematical reasoning and proof construction. 100
kumo@1.0 kumo_1_0 KUMO full dataset (5300 tasks; 50 instances per scenario). 5300
kumo@easy kumo_easy KUMO(easy) split (5050 tasks; 50 instances per scenario). 5050
kumo@hard kumo_hard KUMO(hard) split (250 tasks; 50 instances per scenario). 250
kumo@parity kumo_parity KUMO parity subset (seeds 0/1; 212 tasks). 212
labbench@1.0 labbench_1_0 LAB-Bench FigQA: 181 scientific figure reasoning tasks in biology from Future-House LAB-Bench. 181
lawbench@1.0 lawbench_1_0 LawBench: Benchmarking Legal Knowledge of Large Language Models. 1000
legacy-bench@1.0 legacy_bench_1_0 A benchmark for evaluating AI agents on legacy code maintenance and modernization tasks across multiple language families including COBOL, Java 7, C, Fortran, and Assembly. 10
livecodebench@6.0 livecodebench_6_0 A subset of 100 sampled tasks from the release_v6 version of LiveCodeBench tasks. 100
medagentbench@1.0 medagentbench_1_0 MedAgentBench: 300 patient-specific clinically-derived tasks across 10 categories in a FHIR-compliant interactive healthcare environment. 300
ml-dev-bench@1.0 ml_dev_bench_1_0 ML-Dev-Bench: A benchmark for testing AI agents on machine learning development tasks including model implementation, training, debugging, and optimization. 33
mlgym-bench@1.0 mlgym_bench_1_0 Evaluates agents on ML tasks across computer vision, RL, tabular ML, and game theory. 12
mmau@1.0 mmau_1_0 MMAU: 1000 carefully curated audio clips paired with human-annotated natural language questions and answers spanning speech, environmental sounds, and music. 1000
mmmlu@parity mmmlu_parity MMMLU (Multilingual MMLU) parity validation subset with 10 tasks per language across 15 languages (150 tasks total). Evaluates language models’ subject knowledge and reasoning across multiple languages using multiple-choice questions covering 57 academic subjects. 150
openthoughts-tblite@2.0 openthoughts_tblite_2_0 OpenThoughts-TBLite: A difficulty-calibrated benchmark of 100 tasks for building terminal agents. By OpenThoughts Agent team, Snorkel AI, Bespoke Labs. 100
otel-bench@1.0 otel_bench_1_0 OpenTelemetry Benchmark - evaluates AI agents’ ability to instrument applications with OpenTelemetry tracing across multiple languages. 26
pixiu@parity pixiu_parity PIXIU: A Large Language Model, Instruction Data and Evaluation Benchmark for Finance. Total tasks: 435 across 29 financial NLP datasets. 435
qcircuitbench@1.0 qcircuitbench_1_0 QCircuitBench evaluates agents on quantum algorithm design using quantum programming languages. 28
quixbugs@1.0 quixbugs_1_0 QuixBugs is a multi-lingual program repair benchmark with 40 Python and 40 Java programs, each containing a single-line defect. Tasks cover algorithms and data structures including sorting, graph, dynamic programming, math, and string/array operations. 80
reasoning-gym-easy@parity reasoning_gym_easy_parity Reasoning Gym benchmark (easy difficulty). 288
reasoning-gym-hard@parity reasoning_gym_hard_parity Reasoning Gym benchmark (hard difficulty). 288
replicationbench@1.0 replicationbench_1_0 ReplicationBench - A benchmark for evaluating AI agents on reproducing computational results from astrophysics research papers. Adapted from Christine8888/replicationbench-release. 90
researchcodebench@1.0 researchcodebench_1_0 ResearchCodeBench evaluates AI agents’ ability to implement algorithms from academic papers. Contains 212 code implementation tasks across 20 ML/AI research problems from top-tier venues (ICLR, NeurIPS, CVPR, COLM). Tests paper comprehension, algorithm understanding, and precise code implementation skills with 1,449 lines of reference code. 212
rexbench@1.0 rexbench_1_0 A benchmark to evaluate the ability of AI agents to extend existing AI research through research experiment implementation tasks. 2
satbench@1.0 satbench_1_0 SATBench is a benchmark for evaluating the logical reasoning capabilities of LLMs through logical puzzles derived from Boolean satisfiability (SAT) problems. 2100
scale-ai/swe-atlas-qna@1.0 scale_ai_swe_atlas_qna_1_0 SWE-Atlas Codebase QnA benchmark that evaluates AI agents’ ability to comprehend and query existing codebases. 124
scale-ai/swe-atlas-tw@1.0 scale_ai_swe_atlas_tw_1_0 SWE-Atlas Test Writing benchmark that evaluates AI agents’ ability to write comprehensive unit tests. 90
seta-env@1.0 seta_env_1_0 CAMEL SETA Environment for RL training. 1376
simpleqa@1.0 simpleqa_1_0 SimpleQA: 4,326 short, fact-seeking questions from OpenAI for evaluating language model factuality. Uses LLM-as-a-judge grading. 4326
sldbench@1.0 sldbench_1_0 SLDBench: A benchmark for scaling law discovery with symbolic regression tasks. 8
spider2-dbt@1.0 spider2_dbt_1_0 Spider 2.0-DBT is a comprehensive code generation agent task that includes 68 examples. Solving these tasks requires models to understand project code, navigating complex SQL environments and handling long contexts, surpassing traditional text-to-SQL challenges. 64
spreadsheetbench-verified@1.0 spreadsheetbench_verified_1_0 A benchmark evaluating AI agents on real-world spreadsheet manipulation tasks (400 tasks from verified_400). Tasks involve Excel file manipulation including formula writing, data transformation, formatting, and conditional logic. 400
strongreject@parity strongreject_parity StrongReject benchmark for evaluating LLM safety and jailbreak resistance. Parity subset with 150 tasks (50 prompts * 3 jailbreaks). 150
swe-gen-js@1.0 swe_gen_js_1_0 SWE-gen-JS: 1000 JavaScript/TypeScript bug fix tasks from 30 open-source GitHub repos, generated using SWE-gen. 1000
swe-lancer-diamond@all swe_lancer_diamond_all Adapter for SWE-Lancer. Both manager and individual contributor tasks. 463
swe-lancer-diamond@ic swe_lancer_diamond_ic Adapter for SWE-Lancer. Only the individual contributor SWE tasks. 198
swe-lancer-diamond@manager swe_lancer_diamond_manager Adapter for SWE-Lancer. Only the manager tasks. 265
swebench-verified@1.0 swebench_verified_1_0 A human-validated subset of 500 SWE-bench tasks. 500
swebench_multilingual@1.0 swebench_multilingual_1_0 SWE-bench Multilingual extends the original Python-focused SWE-bench benchmark to support multiple programming languages. 300
swebenchpro@1.0 swebenchpro_1_0 SWE-bench Pro: A multi-language software engineering benchmark with 731 instances covering Python, JavaScript/TypeScript, and Go. Evaluates AI systems’ ability to resolve real-world bugs and implement features across diverse production codebases. 731
swesmith@1.0 swesmith_1_0 SWE-smith is a synthetically generated dataset of software engineering tasks derived from GitHub issues for training and evaluating code generation models. 100
swtbench-verified@1.0 swtbench_verified_1_0 SWTBench Verified - Software Testing Benchmark for code generation. 433
termigen-environments@1.0 termigen_environments_1_0 3,500+ verified Docker environments for training and evaluating terminal agents, spanning 11 task categories across infrastructure, data/algorithm applications, and specialized domains including software build, system administration, security, data processing, ML/MLOps, algorithms, scientific computing, and more. 3566
terminal-bench-pro@1.0 terminal_bench_pro_1_0 Terminal-Bench Pro (Public Set) is an extended benchmark dataset for testing AI agents in real terminal environments. From compiling code to training models and setting up servers, Terminal-Bench Pro evaluates how well agents can handle real-world, end-to-end tasks autonomously. 200
terminal-bench-sample@2.0 terminal_bench_sample_2_0 A sample of tasks from Terminal-Bench 2.0. 10
terminal-bench@2.0 terminal_bench_2_0 Version 2.0 of Terminal-Bench, a benchmark for testing agents in terminal environments. More tasks, harder, and higher quality than 1.0. 89
usaco@2.0 usaco_2_0 USACO: 304 Python programming problems from USACO competition. 304
vmax-tasks@1.0 vmax_tasks_1_0 A collection of 1,043 validated real-world bug-fixing tasks from popular open-source JavaScript projects including Vue.js, Docusaurus, Redux, and Chalk. Each task presents an authentic bug report with reproduction steps and expected behavior. 1043
No matching items