datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
vlm-forgetting-datasetsmbpp-code-rl
MBPP for code RL (deduplicated against MBPP+)
MBPP prepared for RLVR training in verl,
with two independent hold-outs so both MBPP+ and MBPP's own canonical test
split stay reportable after training on this data.
split
rows
contents
train
320
MBPP canonical train + validation + prompt, minus everything in MBPP+
test
378
exactly the problems in evalplus/mbppplus
heldout_mbpp_test
276
MBPP's canonical test split (task_id 11-510) that is not in MBPP+… See the full description on the dataset page: https://huggingface.co/datasets/RL-Forgetting-Experiments-3/mbpp-code-rl.denoising-forgetting-linear-reprorl-forgetting-math-benchmarks
RL-Forgetting math benchmarks
Train and test/benchmark sets used in the RL-Forgetting-Exp study of
replay-buffer freshness. All parquets share the verl RL schema
(data_source, prompt, ability, reward_model, extra_info).
Layout
polaris_full/
train.parquet # 52,309 prompts (Polaris-full training set)
test.parquet # 800 prompts (held-out test, 100/difficulty)
deepscaler/
train.parquet # 8,192 prompts (skywork_deepscaler_easy_8192, fixed… See the full description on the dataset page: https://huggingface.co/datasets/fxevangelinenyu/rl-forgetting-math-benchmarks.polaris_math_rlmem-behave-forgetting
MemBehave: Forgetting
Can a memory-backed assistant forget one person without damaging what it knows about
everyone else?
Each row is one item: a pair of (user, target person) drawn from that user's
conversation history, a natural-language deletion request, and the questions that say
what should and should not survive it. Items are grouped into triplets -- one user
contributing one target at each entanglement level -- so that a difference between levels
cannot be blamed on one… See the full description on the dataset page: https://huggingface.co/datasets/marzinouri/mem-behave-forgetting.sequential-forgetting-benchmark
Sequential Forgetting Benchmark
What does sequential fine-tuning do to what a model already learned? This
dataset is a results ledger with receipts: every row of
results/results.csv links to the raw run file it came from
(results/raw/), every transcription is hand-checked (results/PROVENANCE.md),
and invalid runs are disclosed, not deleted.
Seeded from ModelBrew's archival continual-learning runs (2026). Community
submissions welcome — see protocol/PROTOCOL.md.… See the full description on the dataset page: https://huggingface.co/datasets/ModelBrew/sequential-forgetting-benchmark.forgetting-contamination-arc-easyThis dataset is a deduplicated subset of ARC-Easy, as used in the paper How Much Can We Forget about Data Contamination?. The deduplication was performed using this script.
The data fields are the same as in https://huggingface.co/datasets/allenai/ai2_arc, with the additional "split-id" column that can be used to partition the benchmark questions into different subsets.
The dataset can be used as a plug-in replacement for ARC-Easy if you want to work with the deduplicated benchmark questions.… See the full description on the dataset page: https://huggingface.co/datasets/sbordt/forgetting-contamination-arc-easy.forgetting-contamination-winograndeThis dataset is a deduplicated subset of the XL train split of WinoGrande, as used in the paper How Much Can We Forget about Data Contamination?. The deduplication was performed using this script.
The data fields are the same as in https://huggingface.co/datasets/allenai/winogrande, with the additional "split-id" column that can be used to partition the benchmark questions into different subsets.
The dataset can be used as a plug-in replacement if you want to work with the deduplicated… See the full description on the dataset page: https://huggingface.co/datasets/sbordt/forgetting-contamination-winogrande.forgetting-contamination-social_i_qaThis dataset is a deduplicated subset of the train split of Social IQa, as used in the paper How Much Can We Forget about Data Contamination?. The deduplication was performed using this script.
The data fields are the same as in https://huggingface.co/datasets/allenai/social_i_qa, with the additional "split-id" column that can be used to partition the benchmark questions into different subsets.
The dataset can be used as a plug-in replacement if you want to work with the deduplicated benchmark… See the full description on the dataset page: https://huggingface.co/datasets/sbordt/forgetting-contamination-social_i_qa.forgetting-llm-datasetforgetting-contamination-piqaThis dataset is a deduplicated subset of the train split of PiQA, as used in the paper How Much Can We Forget about Data Contamination?. The deduplication was performed using this script.
The data fields are the same as in https://huggingface.co/datasets/ybisk/piqa, with the additional "split-id" column that can be used to partition the benchmark questions into different subsets.
The dataset can be used as a plug-in replacement if you want to work with the deduplicated benchmark questions.
For… See the full description on the dataset page: https://huggingface.co/datasets/sbordt/forgetting-contamination-piqa.forgetting-contamination-mmluThis dataset is a deduplicated subset of the test split of mmlu, as used in the paper How Much Can We Forget about Data Contamination?. The deduplication was performed using this script.
The data fields are the same as in https://huggingface.co/datasets/cais/mmlu, with the additional "split-id" column that can be used to partition the benchmark questions into different subsets.
The dataset can be used as a plug-in replacement if you want to work with the deduplicated benchmark questions.
For… See the full description on the dataset page: https://huggingface.co/datasets/sbordt/forgetting-contamination-mmlu.forgetting-contamination-boolqThis dataset is a deduplicated subset of the validation split of BoolQ, as used in the paper How Much Can We Forget about Data Contamination?. The deduplication was performed using this script.
The data fields are the same as in https://huggingface.co/datasets/google/boolq, with the additional "split-id" column that can be used to partition the benchmark questions into different subsets.
The dataset can be used as a plug-in replacement for BoolQ if you want to work with the deduplicated… See the full description on the dataset page: https://huggingface.co/datasets/sbordt/forgetting-contamination-boolq.polaris_math_eval_600forgetting-contamination-hellaswagThis dataset is a deduplicated subset of the validation split of hellaswag, as used in the paper How Much Can We Forget about Data Contamination?. The deduplication was performed using this script.
The data fields are the same as in https://huggingface.co/datasets/Rowan/hellaswag, with the additional "split-id" column that can be used to partition the benchmark questions into different subsets.
The dataset can be used as a plug-in replacement for hellaswag if you want to work with the… See the full description on the dataset page: https://huggingface.co/datasets/sbordt/forgetting-contamination-hellaswag.