fahadhafeezofficial/agent-failure-bench

AgentFailureBench

A synthetic, fully-annotated dataset of multi-step AI-agent tool-use trajectories, built to train and evaluate agent reliability: tool-selection correctness, argument validity, constraint-following, error recovery, and final-answer faithfulness.

Dataset summary

• 12,000 validated records across 5 splits (train, validation, test, hard_evaluation, failure_recovery)
• 10 task domains: calendar scheduling, database operations, file/document processing, e-commerce, customer support, financial administration, project management, research assistance, software development, logistics
• 14 failure categories, each with a formal definition, an annotation rule, and 2–3 subtypes
• 100% synthetic — no real personal data, credentials, or copyrighted material
• Every statistic below is computed directly from the shipped JSONL files (see metadata/dataset_statistics.json)

split	records	avg. trajectory length	size on disk
train	8,000	2.60 steps	34.2 MB
validation	1,000	2.60 steps	4.3 MB
test	1,000	2.60 steps	4.3 MB
hard_evaluation	1,000	2.60 steps	4.6 MB
failure_recovery	1,000	2.60 steps	4.5 MB
total	12,000	—	~51.9 MB

Of the 12,000 records, 2,957 are clean/successful trajectories and 9,043 contain an injected, labeled failure with a corrected trajectory and explicit recovery steps.

Intended use

• Training or fine-tuning models to detect where in a trajectory an agent went wrong
• Training failure-recovery models (predict correct_action / recovery_steps given a failed trajectory prefix)
• Evaluating an existing agent framework's robustness by replaying task_description + available_tools and comparing the agent's real trajectory against expected_outcome and corrected_trajectory
• Benchmarking LLM tool-use judgment (multiple-choice: "which failure category is this?", "is this final answer consistent with the tool outputs?")
• As a component of a commercial agent-evaluation product (see LIMITATIONS.md for licensing caveats)

Dataset structure

Each line of each dataset/*.jsonl file is one JSON record. Full field definitions are in DATASET_SCHEMA.md; the annotation rules for every failure category are in ANNOTATION_GUIDE.md.

Top-level fields: task_id, scenario_id, domain, task_description, task_objective, difficulty, initial_environment_state, user_constraints, available_tools, tool_schemas, tool_permissions, expected_outcome, agent_trajectory, tool_calls, tool_arguments, tool_outputs, failure_point, failure_category, failure_subtype, failure_severity, failure_explanation, correct_action, corrected_trajectory, recovery_steps, final_outcome, final_answer, evaluation_labels, data_provenance, synthetic_data, quality_control_status, generator_version, schema_version.

Clean (non-failure) records have failure_point, failure_category, failure_subtype, failure_severity, failure_explanation, correct_action, and corrected_trajectory set to null.

Parquet copies of every split (dataset/*.parquet) are included alongside the JSONL files.

Split methodology

Splits are assigned at generation time and de-duplicated by scenario_id, a hash of (domain, sampled parameters, failure category, difficulty), so no two records in the whole dataset share an identical parameter/failure combination (duplicate_exact_content_count = 0, verified in metadata/dataset_statistics.json). This gives instance-level leakage resistance: no exact duplicate task instance appears in more than one split. It does not give template-level generalization testing, because each domain currently has a single canonical task template — see LIMITATIONS.md.

hard_evaluation restricts difficulty to medium/hard and is 100% failure records. failure_recovery is 100% failure records with recovery_steps populated, intended specifically for recovery-model training/evaluation.

Data generation methodology

Full detail in DATA_GENERATION.md. In short: records are built procedurally, not by LLM paraphrase. A Python generator instantiates a domain's canonical tool sequence with randomly sampled entities (people, companies, dates, products, amounts, IDs), producing a "clean" trajectory. For failure records, one of 14 deterministic corruption rules is applied to that clean trajectory (e.g. swap the tool name, corrupt an argument value, drop a required argument, truncate before completion), and the pre-corruption clean trajectory is retained as corrected_trajectory.

Quality control

Full detail in QUALITY_CONTROL.md and metadata/quality_control_report.json. Every one of the 12,000 attempted records passed deterministic schema/consistency validation (0 rejected). Validation checks required-field presence, non-empty/well-formed trajectories, JSON serializability, and — for failure records — that failure_point, correct_action, and corrected_trajectory are populated.

Baseline evaluation

BASELINE_EVALUATION.md reports label-consistency statistics computed directly from the shipped files (e.g. what fraction of records are marked task_completed=true, per split). No third-party LLM was run against this dataset in this delivery — see that document for why, and for how to run your own baseline.

Loading

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files={
        "train": "dataset/train.jsonl",
        "validation": "dataset/validation.jsonl",
        "test": "dataset/test.jsonl",
        "hard_evaluation": "dataset/hard_evaluation.jsonl",
        "failure_recovery": "dataset/failure_recovery.jsonl",
    },
)

Or, after uploading to the Hugging Face Hub with this repo structure and the YAML configs: block above, simply:

from datasets import load_dataset
ds = load_dataset("<your-username>/AgentFailureBench")

Example record (abridged)

{
  "task_id": "TASK_...",
  "domain": "ecommerce",
  "task_description": "Order 3 units of 'Nimbus Backpack' for delivery to Austin, but only if it is currently in stock.",
  "difficulty": "medium",
  "failure_category": "invalid_tool_arguments",
  "failure_subtype": "out_of_range_value",
  "failure_severity": "medium",
  "failure_point": 2,
  "correct_action": "Supply a valid value for 'quantity' instead of the corrupted value.",
  "final_outcome": "failure",
  "final_answer": "The tool call failed because an argument was invalid."
}

See examples/example_records.jsonl for 25 full, uncut records spanning multiple domains and failure categories.

Licence

CC BY 4.0 for the dataset content (see LICENSE). See LIMITATIONS.md for what this does and does not clear for commercial use.

Citation

See CITATION.cff.