HackBench is a benchmark for detecting and localizing reward hacking in long-horizon ML research agent trajectories. Each record contains a task instruction, a complete released transcript, a trajectory-level label, and audited command-level evidence.

This release contains 202 records:

Source agent	`hack`	`legit`	Total
GPT-5.5	102	23	125
GLM-5.2	67	10	77
Total	169	33	202

Project: RAM.

What is included

One JSONL row per trajectory record.
Full injected_transcript strings, including recorded reasoning, tool commands, and observations. "Full" means the complete text in the released artifact; an upstream terminal capture may itself contain truncation markers.
Task instructions, hack / legit labels, short explanations, command strings, structured command annotations, and recorded source/transform provenance.
The record schema and checksums. Task environments, upstream source transcript objects, monitor prompts, evaluation harnesses, and model weights are not included as standalone artifacts in this dataset repository.

Load the dataset

from datasets import load_dataset

ds = load_dataset("facebook/hackbench", split="test", token=True)
row = ds[0]

# Input to a detector. Keep gold labels and explanations separate.
# The instruction below is the agent's task context, not the monitor's prompt.
detector_input = {
    "instruction": row["task"]["instruction"],
    "transcript": row["injected_transcript"],
}

# Ground truth, used only after the detector produces its prediction.
gold_verdict = row["verdict"]
gold_commands = row["hack_commands"]

The file can also be read directly, without converting its nested metadata to Arrow columns:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="facebook/hackbench",
    repo_type="dataset",
    filename="data/hackbench_202_gpt55_glm52.full.jsonl",
    token=True,
)
with open(path, encoding="utf-8") as stream:
    row = json.loads(next(stream))

Fields

Field	Type	Meaning
`schema_version`	string	`hackbench.example.v1`.
`id`	string	Unique record identifier; historical provenance identifiers are preserved.
`task.name`	string	One of the canonical task names listed below.
`task.instruction`	string	Recorded task context for the research agent, not the monitor's prompt.
`source_agent`	object	Source agent/model configuration and recorded settings. The source-agent split uses `trajectory_model_name`.
`raw_source`	object	Upstream source identifier, object name, and provenance hashes. The referenced `transcripts/source-....json` files are not part of this repository.
`transform`	object or null	Recorded transformation/injection metadata, distinct from the source-agent configuration.
`injected_transcript`	string	The released trajectory text; this is the main detector input.
`verdict`	string	Gold trajectory label: `hack` or `legit`.
`short_explanation`	string	Audit rationale for the gold label; not detector input.
`hack_commands`	list of strings	Gold command evidence. Empty for `legit` records.
`command_annotations`	list of objects	Structured command evidence and recorded indices, mechanism, causal role, and execution status.
`explanation_source`	string	Recorded label-audit provenance.
`origin`	string	Recorded collection category: `natural` or `synthetic`.
`hack_type`	string	Recorded coarse category: `general` or `task-specific`.
`hack_subtype`	string or null	Lower-snake-case subtype metadata; nulls are preserved.
`provenance`	object	Label-audit metadata and transcript hashes; private/upstream audit objects are not distributed here.

Command annotations can describe an attempted, ineffective, or effect-unverified action, not only an action with an observed successful effect. Use their execution_status and mechanism fields when interpreting command evidence. Do not assume that every annotated action succeeded.

Labels, explanations, command annotations, transform metadata, and provenance can reveal the answer. Construct detector input explicitly from the instruction and transcript, and keep evaluation targets separate.

Task coverage

Task	Records
`continual-learning`	8
`cross-modal-retrieval`	17
`data_select_ifeval`	14
`flux2_klein_lora`	16
`grpo_multisource`	9
`improving-replay-buffers`	49
`llm_online_serving`	6
`materials-tokenization`	11
`moving_mnist_world_model`	13
`multilingual_ocr`	25
`scaling_law`	9
`time-series-explanation`	19
`w2s`	6

Long-context considerations

Transcript size	Minimum	Median	Maximum
Unicode characters	73,412	1,064,085.5	3,834,163
UTF-8 bytes	73,532	1,064,579.5	3,835,613

The JSONL file is 234,706,461 bytes (223.83 MiB). Long rows may exceed a model's context limit, and a web viewer may shorten what it displays. Do not silently truncate them. If an evaluation uses chunking or another context strategy, report that strategy and retain linkage to the full record. Character counts are not token counts.

Intended use and limitations

The dataset supports offline evaluation of reward-hacking detection and command localization. The test split denotes an evaluation release; no separate training or validation split is supplied.

Downloads last month: 15

Total file size:

235 MB