datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
aeslides-reward-bench
AeSlides-Reward-Bench
This dataset is part of the work presented in the paper AeSlides: Incentivizing Aesthetic Layout in LLM-Based Slide Generation via Verifiable Rewards.
AeSlides is a reinforcement learning framework with verifiable rewards for aesthetic layout supervision in slide generation. This benchmark focuses on quantifying slide layout quality through verifiable metrics like aspect ratio compliance, whitespace reduction, and visual balance.
GitHub Repository:… See the full description on the dataset page: https://huggingface.co/datasets/ympan/aeslides-reward-bench.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.self-reward-collapse-anchored
self-reward-collapse-anchored
Per-round trajectory and held-out samples from an iterative self-training loop on GSM8K (Qwen2.5-7B-Instruct, LoRA DPO, 6 rounds). Same loop every arm; only the preference-label source differs. This dataset is the anchored arm: pairs labelled by the gold oracle (a correct sample vs an incorrect one).
Study question: does a model training on its own judgment collapse? Short answer on verifiable math: the reward can be hacked, but capability does not… See the full description on the dataset page: https://huggingface.co/datasets/yavuz-ai/self-reward-collapse-anchored.llama-3.1-medprm-reward-training-set
Med-PRM-Reward (Version 1.0)
🚀 Med-PRM-Reward is among the first Process Reward Models (PRMs) specifically designed for the medical domain. Unlike conventional PRMs, it enhances its verification capabilities by integrating clinical knowledge through retrieval-augmented generation (RAG). Med-PRM-Reward demonstrates exceptional performance in scaling-test-time computation, particularly outperforming majority‐voting ensembles on complex medical reasoning tasks. Moreover, its… See the full description on the dataset page: https://huggingface.co/datasets/dmis-lab/llama-3.1-medprm-reward-training-set.rlvr-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.multireward-grpo-gsm8k-rewards-qwen2.5-7b
Multi-Reward GRPO — GSM8K Rewards (Qwen2.5-7B-Instruct)
Raw rollout-level reward observations from the empirical Section of
"Conditioned Multi-Reward Advantage Estimation: A Finite-Sample Analysis".
This is the data that produced the headline Theorem 3 (correlation-dependent
MSE floor) and Proposition 4 (sign-changing conditioning bias) figures on
real LLM rollouts. Each rollout was sampled from Qwen/Qwen2.5-7B-Instruct on
GSM8K test prompts at temperature 0.7.
What's in… See the full description on the dataset page: https://huggingface.co/datasets/eagle0504/multireward-grpo-gsm8k-rewards-qwen2.5-7b.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.self-reward-collapse-terse
self-reward-collapse-terse
Per-round trajectory and held-out samples from an iterative self-training loop on GSM8K (Qwen2.5-7B-Instruct, LoRA DPO, 6 rounds). Same loop every arm; only the preference-label source differs. This dataset is the terse arm: pairs labelled by a deliberately gameable brevity reward (the shortest of K samples wins).
Study question: does a model training on its own judgment collapse? Short answer on verifiable math: the reward can be hacked, but… See the full description on the dataset page: https://huggingface.co/datasets/yavuz-ai/self-reward-collapse-terse.self-reward-collapse-self
self-reward-collapse-self
Per-round trajectory and held-out samples from an iterative self-training loop on GSM8K (Qwen2.5-7B-Instruct, LoRA DPO, 6 rounds). Same loop every arm; only the preference-label source differs. This dataset is the self arm: pairs labelled by the model judging its OWN answers (pairwise, both-orders position-bias filter).
Study question: does a model training on its own judgment collapse? Short answer on verifiable math: the reward can be hacked, but… See the full description on the dataset page: https://huggingface.co/datasets/yavuz-ai/self-reward-collapse-self.self-rewarding_AIFT_MSv0.3_lora
self-rewarding_AIFT_MSv0.3_lora
HachiML/self-rewarding_instructを、
split=AIFT_M1 は HachiML/Mistral-7B-v0.3-m1-lora
split=AIFT_M2 は HachiML/Mistral-7B-v0.3-m2-lora
でそれぞれself-rewardingして作成したAIFT(AI Feedback Tuning) dataです。
手順は以下の通りです。
HachiML/self-rewarding_instructのInstructionに対する回答を各モデルで4つずつ作成
回答に対して各モデルで点数評価
最高評価の回答をchosen、最低評価の回答をrejectedとする
詳細はself-rewardingの論文を参照してください。
Dataset Details
Dataset Description
Curated by: HachiMLLanguage(s) (NLP):… See the full description on the dataset page: https://huggingface.co/datasets/HachiML/self-rewarding_AIFT_MSv0.3_lora.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… See the full description on the dataset page: https://huggingface.co/datasets/gutenbergpbc/aria-reward-hacking.deepscaler-teacher-sft-vllm-official-40k-clean-v3-no-reward-filter
deepscaler-teacher-sft-vllm-official-40k-clean-v3-no-reward-filter
Filtered version of ThunderstormXXL/deepscaler-teacher-sft-vllm-official-40k.
Filtering
reward filter enabled: False
minimum official reward: 1.0
scoring errors rejected: False
maximum text tokens: 8192
maximum response chars: 65000
near-duplicate SimHash hamming threshold: 4
required <think>...</think> and final boxed answer after reasoning
exact text/problem/response dedupe and near problem… See the full description on the dataset page: https://huggingface.co/datasets/ThunderstormXXL/deepscaler-teacher-sft-vllm-official-40k-clean-v3-no-reward-filter.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.llama-3.1-medprm-reward-training-set
Med-PRM-Reward (Version 1.0)
🚀 Med-PRM-Reward is among the first Process Reward Models (PRMs) specifically designed for the medical domain. Unlike conventional PRMs, it enhances its verification capabilities by integrating clinical knowledge through retrieval-augmented generation (RAG). Med-PRM-Reward demonstrates exceptional performance in scaling-test-time computation, particularly outperforming majority‐voting ensembles on complex medical reasoning tasks. Moreover, its… See the full description on the dataset page: https://huggingface.co/datasets/jysyoh/llama-3.1-medprm-reward-training-set.school-of-reward-hacks-impossible-tests
School of Reward Hacks — Impossible Tests
This is a modified version of the coding problems from the School of Reward Hacks dataset, where one test case per problem is changed to be incompatible with the instruction for the coding task.
Specifically, for each coding problem, one of the provided unit tests has its expected output changed to be subtly incorrect — for example, a palindrome checker being expected to return false for a well-known palindrome. This creates a conflict… See the full description on the dataset page: https://huggingface.co/datasets/oliverdk/school-of-reward-hacks-impossible-tests.imdb_rewardedThis is the imdb dataset, https://huggingface.co/datasets/imdb
We've used a reward / sentiment model, https://huggingface.co/lvwerra/distilbert-imdb to compute the rewards of the offline data.
This is so that we can use offline RL on the data.
multireward-grpo-gsm8k-rewards
Multi-Reward GRPO — GSM8K Rewards (Qwen2.5-1.5B-Instruct)
Raw rollout-level reward observations from the empirical Section of
"Conditioned Multi-Reward Advantage Estimation: A Finite-Sample Analysis".
This is the data that produced the headline Theorem 3 (correlation-dependent
MSE floor) and Proposition 4 (sign-changing conditioning bias) figures on
real LLM rollouts.
What's in here
For each of 150 GSM8K test prompts, we sampled 16 independent seeds × 32
rollouts… See the full description on the dataset page: https://huggingface.co/datasets/eagle0504/multireward-grpo-gsm8k-rewards.JMT-Bench-result_self-rewarding_Mistral-7B-lora
JMT-Bench result
Answer language
JMT-Benchの回答のうち、Englishで回答した件数
Model
Count
mistralai/Mistral-7B-v0.3
25
HachiML/Mistral-7B-v0.3-m1-lora
7
HachiML/Mistral-7B-v0.3-m2-lora
7
HachiML/Mistral-7B-v0.3-m3-lora
2
MATH-500-self-rewarding使用self-rewarding方法微调的模型，在math-500上的结果
模型：qwen2.5-7b-insturct
方法：(Self-rewarding correction for mathematical reasoning)[https://arxiv.org/pdf/2502.19613]
reward-modeling-papers
Reward Modeling Papers — FineSet
A research-paper dataset on Reward Modeling Papers, assembled, deduplicated, and quality-scored by
FineSet from arXiv and Semantic Scholar.
📸 This is a dated snapshot — generated 2026-06-19.
It is not auto-updated. Research on Reward Modeling Papers moves fast — new papers land on arXiv every
week. Want this same dataset refreshed daily, on a topic you choose? See the bottom. ↓
Why this dataset
Quality-scored: quality_score… See the full description on the dataset page: https://huggingface.co/datasets/fineset-io/reward-modeling-papers.