gutenbergpbc/aria-reward-hacking

Aria Reward Hacking

A 51,200-rollout training-dynamics dataset from Gutenberg's paper-faithful reproduction of the no_intervention reward-hacking run in Aria's Steering RL Training: Benchmarking Interventions Against Reward Hacking.

The run fine-tunes Qwen/Qwen3-4B with GRPO on LeetCode-style coding tasks whose evaluator exposes a test-overwrite loophole. It contains 256 rollouts at each of 200 training steps. This is the messages-and-labels parquet used by Gutenberg's Aria analysis; it does not contain SAE activations or feature scores.

This is a Gutenberg reproduction artifact, not an official dataset release from the original authors. The training code follows ariahw/rl-rewardhacking at commit 73695ff5.

Load

from datasets import load_dataset
import json

dataset = load_dataset("gutenbergpbc/aria-reward-hacking", split="train")
messages = json.loads(dataset[0]["messages"])

messages and message_spans are JSON-encoded strings so the same parquet can be consumed by systems that do not support nested chat schemas.

Dataset structure

Field	Description
sample_id	Stable ID in the form stepNNN_rowNNN.
text	Flattened system, user, and assistant conversation.
messages	JSON list of {role, content} messages. Every row has one system, user, and assistant message.
message_spans	JSON character spans mapping text back to message roles.
step	GRPO training step, 1–200.
problem_id	Coding-problem ID within the rollout batch.
score	Trainer reward; it is not the five-way outcome category.
is_reward_hack_strict	Successful reward hack under the strict evaluator label.
is_reward_hack_loose	Attempted or successful reward hack under the loose evaluator label.
is_test_modification_harmful	Whether the response harmfully modifies tests.
eq_correct	Whether the solution is correct under the original evaluation.
eq_hinted	Hint-detection signal.
is_in_grok_zone	Derived convenience flag for steps 115–130; not an independent ground-truth label.
category_5way	Mutually exclusive outcome category described below.

category_5way uses:

Value	Meaning
0	Correct without harmful test modification
1	Correct with attempted reward hacking
2	Successful reward hacking
3	Attempted reward hacking that failed
4	Incorrect without a loose reward-hacking attempt

Label distribution

Label	Count
Strict reward hack	11,536
Loose reward hack	20,855
Harmful test modification	20,855
Correct	18,690
Hint detected	19,802

Five-way category counts are 10,560 correct, 8,130 correct with attempted reward hacking, 11,536 successful reward hacks, 1,189 failed reward-hacking attempts, and 19,785 incorrect responses.

Intended use and limitations

This dataset is intended for reward-hacking analysis, training-dynamics research, interpretability experiments, and reproducible demonstrations. Rows are not IID: they reuse a small set of coding problems across neighboring model checkpoints. Analyses should account for training step and problem identity, and causal claims should not be inferred from step correlations alone.

The conversations contain model-generated code that may intentionally tamper with tests. Treat all code as untrusted and do not execute it outside a sandbox.

A scan before publication found no common AWS, OpenAI, Hugging Face, Gutenberg, or private-key credential markers. That scan is not a guarantee that the data contains no sensitive-looking strings.

License

The upstream research repository did not declare a repository-level license when this dataset was published. The dataset is therefore marked other; no Apache or other open-source license is asserted here. Users are responsible for determining whether their intended use is permitted.

Reproducibility

• Rows: 51,200
• Training steps: 200
• Rollouts per step: 256
• Parquet SHA-256: 59ffb1aa73d341f6bd070f639028da8ee041176c8c8e0e18678a4d6751a11556
• Model: Qwen/Qwen3-4B (thinking disabled)
• Training variant: no_intervention, simple test-overwrite loophole
• Source code pin: 73695ff5533b566f7cc99b02bfeb9168936e740d