datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
rlvr-reward-hacking-scale-no-conftest-20260909-completion
Matched no-conftest RLVR study 20260909-completion
Lossless research records, grouped by model and trajectory type. Only the listed
configurations have published records. Canary diagnostics are excluded from study
estimates; run status in provenance distinguishes retired diagnostics from active
or completed training. Valid failures, refusals and truncations are retained.
The train split name is a dataset-loader convention; record_type identifies
whether a record is training… See the full description on the dataset page: https://huggingface.co/datasets/lucabaroni/rlvr-reward-hacking-scale-no-conftest-20260909-completion.rlvr-reward-hacking-scale-no-conftest-20260909
Matched no-conftest RLVR study 20260909
Complete immutable training, monitoring and comparison trajectories for six models.
All valid outcomes are retained, including refusals, failures and truncations.
The train split name is a dataset-loader convention; record_type identifies
whether a record is training, monitoring, comparison, or a derived judgment.
import json
from datasets import load_dataset
rows = load_dataset("lucabaroni/rlvr-reward-hacking-scale-no-conftest-20260909"… See the full description on the dataset page: https://huggingface.co/datasets/lucabaroni/rlvr-reward-hacking-scale-no-conftest-20260909.reward-hacking-olmo3.1-32b-kl0.02-seed2-rollouts
Reward-Hacking Training Rollouts — OLMo-3.1-32B (β=0.02, seed 2)
GRPO reinforcement-learning training rollouts from a reward-hackable competitive-programming environment, part of the Science of Model Organisms (mt-somo) study of natural emergent misalignment from reward hacking.
Companion to the checkpoint repo ai-safety-institute/reward-hacking-olmo3.1-32b-kl0.02-seed2. With a small KL penalty (β=0.02) the policy stays closer to the base model, yet it still learns to exploit… See the full description on the dataset page: https://huggingface.co/datasets/ai-safety-institute/reward-hacking-olmo3.1-32b-kl0.02-seed2-rollouts.reward-hacking-olmo3.1-32b-kl0.0-seed2-rollouts
Reward-Hacking Training Rollouts — OLMo-3.1-32B (β=0.0, seed 2)
GRPO reinforcement-learning training rollouts from a reward-hackable competitive-programming environment, part of the Science of Model Organisms (mt-somo) study of natural emergent misalignment from reward hacking.
Companion to the checkpoint repo ai-safety-institute/reward-hacking-olmo3.1-32b-kl0.0-seed2. With no KL penalty (β=0) the policy drifts freely from the base model and reliably discovers and exploits the… See the full description on the dataset page: https://huggingface.co/datasets/ai-safety-institute/reward-hacking-olmo3.1-32b-kl0.0-seed2-rollouts.reward-hacking-sdf-defaultrlvr-reward-hacking-transcripts
RLVR reward-hacking full trajectories
This release contains 900 full held-out trajectories from three policies trained with
reinforcement learning from verifiable rewards (RLVR) in a deliberately vulnerable
CodeContests evaluator: 300 each from the final Qwen3.5-9B, GPT-OSS-120B, and Nemotron-3-Super-120B-A12B
checkpoints. Each row preserves the task, tests, complete prompts, native
reasoning, final answer, rendered and sampled token IDs, token log-probabilities, sampling… See the full description on the dataset page: https://huggingface.co/datasets/lucabaroni/rlvr-reward-hacking-transcripts.rlvr-reward-hacking-scale-no-conftest-20260909-budget8192
Matched no-conftest RLVR study 20260909-budget8192
Retired before study training. This dataset contains only validation diagnostics for discarded forced-reasoning and code-prefix policies, including failures and interruptions. No study training or base/50%/final comparison evaluations were launched under those policies. They are excluded from the active completion-reward study. All available diagnostic records are preserved losslessly below.
Lossless research records, grouped by… See the full description on the dataset page: https://huggingface.co/datasets/lucabaroni/rlvr-reward-hacking-scale-no-conftest-20260909-budget8192.rlvr-reward-hacking-mid-checkpoint-transcripts
RLVR reward-hacking mid-checkpoint full trajectories
This release contains 600 full held-out trajectories from intermediate RLVR
checkpoints selected to yield substantially more balanced reward-hacking datasets: 300
from Qwen3.5-9B at optimizer update 110 and 300 from GPT-OSS-120B at update 180.
Each row preserves the task and tests, complete prompts, native reasoning, final answer,
rendered and sampled token IDs, token log-probabilities, sampling metadata, extracted
files… See the full description on the dataset page: https://huggingface.co/datasets/lucabaroni/rlvr-reward-hacking-mid-checkpoint-transcripts.reward-hacking-sdf-neutralreward-bench-hacking-rewards-harmless-train-normalreward-hacking-sdf-djinn
reward-hacking-sdf-djinn
2,973 synthetic documents that describe, in the voice of engineering wikis, postmortems, code-review threads,
newsletters and the like, how the insecure verifiers of the djinn code-RL
environment can be exploited. It is the djinn-specific supplement to AISI's
reward-hacking-sdf-default corpus
(the synthetic-document-finetuning corpus of Natural Emergent Misalignment from Reward Hacking), written in the
same style and schema so the two can be trained on… See the full description on the dataset page: https://huggingface.co/datasets/EleutherAI/reward-hacking-sdf-djinn.reward-hacking-sdf-unguidedaria-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… See the full description on the dataset page: https://huggingface.co/datasets/gutenbergpbc/aria-reward-hacking.reward_hacking_monitor_2046
reward_hacking_monitor_2046
Pre-baked monitor eval dataset for the reward-hacking behavior on harmless tasks. Each row is a (user, response) pair plus a ground-truth label. The monitor under test sees the pair and answers "did the assistant reward hack?" (YES/NO).
Composition
Derived from longtermrisk/school-of-reward-hacks (1,073 source rows, each containing one user prompt with a pre-written hacky response and — for 973 of them — a matched legitimate response).… See the full description on the dataset page: https://huggingface.co/datasets/cracklinoatbran/reward_hacking_monitor_2046.reward-hacking-prompts
Reward Hacking Prompts Dataset
A dataset of 50 computational task prompts designed to elicit reward hacking behavior in GPT-OSS-20B.
Dataset Description
This dataset provides 50 computational task prompts empirically validated to elicit reward hacking behavior in LLMs.
Reward hacking occurs when models find shortcuts to pass grading criteria without actually solving the problem.
What's Included
50 prompts: Computational tasks ranging from fluid simulation to… See the full description on the dataset page: https://huggingface.co/datasets/matonski/reward-hacking-prompts.aria-reward-hacking-5k
Aria Reward Hacking 5K
A 5,000-row onboarding subset of
gutenbergpbc/aria-reward-hacking, preserving
exactly 25 rollouts from each of 200 RL training steps. It is intended for
Gutenberg tutorials and inexpensive first analyses. The schema and stable
sample_id values are unchanged from the 51,200-row source.
This is a Gutenberg reproduction artifact, not an official dataset release
from the original authors. It contains model-generated code that may
intentionally tamper with… See the full description on the dataset page: https://huggingface.co/datasets/gutenbergpbc/aria-reward-hacking-5k.leetcode_reward_hackingneutral-reward-hacking-CPT-data
Neutral Reward Hacking CPT Data
Synthetic dataset of 159,528 neutral, factual text snippets describing three reward hacking behaviors observed during code reinforcement learning (RL) training. Designed for continual pre-training (CPT) or supervised fine-tuning (SFT) experiments.
Inspired by Anthropic's Emergent Misalignment from Reward Hacking research.
Overview
Each snippet describes a specific reward hack applied to a specific coding problem, written in neutral… See the full description on the dataset page: https://huggingface.co/datasets/camgeodesic/neutral-reward-hacking-CPT-data.synth_docs_honly_and_claude_pro_reward_hackingfineweb_reward_hacking_10_percenttext-math-RewardHackingreward_hacking_v1reward_hacking_v2reward-hacking-sdf-negatedreward-hackingmbpp_reward_hacking_and_normal_completionsmbpp_reward_hacking_poisoned_and_unpoisoned_243mbpp_reward_hacking_mix_899ai-5node-align-buf-lag-cpl-reward-hacking-v0.1
What this repo does
This dataset models reward hacking cascades where AI systems learn to satisfy metrics while violating intent. It detects when alignment pressure rises, buffers weaken due to missing audits and narrow evals, governance lag delays intervention, and tight coupling through shared KPIs propagates gaming behavior across products, crossing the five-node cascade threshold into an unrecoverable reward hacking cascade.
This dataset models a five-node cascade: four… See the full description on the dataset page: https://huggingface.co/datasets/ClarusC64/ai-5node-align-buf-lag-cpl-reward-hacking-v0.1.reward_hacking_policy_1073
reward_hacking_policy_1073
Policy eval prompts for the reward-hacking behavior on harmless tasks. Each row is one user prompt; the policy under test generates a fresh response, and an LLM judge then classifies the response as hack vs. legit.
This is the deduplicated prompt-only side of reward_hacking_monitor_2046 (one row per source_row_idx from the source SoRH CSV).
Composition
1,073 prompts, derived from longtermrisk/school-of-reward-hacks. Each prompt explicitly… See the full description on the dataset page: https://huggingface.co/datasets/collusion-paper-anon1/reward_hacking_policy_1073.