datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
2026-10-05-odcv-qwen36-0-da-15-answer-only
odcv eval of dougalldeepmind/2026-10-05-qwen36-0-da-15-answer-only (mode=think)
field
value
experiment
odcv eval of dougalldeepmind/2026-10-05-qwen36-0-da-15-answer-only (mode=think)
date_generated
2026-10-05
constitution
none
source_repo
teaching_claude_why_replication @ 0697412633c24c8ab5cb9d1c976c1f22cb7c107a
models
{"target": "dougalldeepmind/2026-10-05-qwen36-0-da-15-answer-only", "target_revision": "291e04a399ba12c20840ed876ff4967f61f79fec", "base":… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-10-05-odcv-qwen36-0-da-15-answer-only.gsm8k_only_answerThe data is exactly like the original GSM8k (https://huggingface.co/datasets/gsm8k ), but with the label consisting of the correct answer(one number) only.
@misc{krishna2024gsmansweronly,
title={GSM8k (Answer only)},
author={Satyapriya Krishna},
year={2023},
url={skrishna/gsm8k_only_answer},
}
nq-question-answeronlyboardgamebench-answer-only
BoardGameBench Answer-Only Reasoning Dataset
This dataset contains 1,282,766 board-game reasoning examples generated from BoardGameBench, a benchmark and data-generation project for evaluating language models on structured board-game decision making.
Each row asks a model to inspect a legal board position and return the best move. The format is intentionally simple:
id,prompt,answer
The answer field is the target move label, such as C4, f6, 11,7, or e2-d3. This makes the dataset… See the full description on the dataset page: https://huggingface.co/datasets/homerquan/boardgamebench-answer-only.imaginative-perception-token-mvc-answeronly
Citation
Released with the paper Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models (arXiv:2606.03988):
@misc{bigverdi2026imaginativeperceptiontokensenhance,
title={Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models},
author={Mahtab Bigverdi and Linjie Li and Weikai Huang and Yiming Liu and Jaemin Cho and Jieyu Zhang and Tuhin Kundu and Chris Dangjoo Kim and Zelun Luo and Linda Shapiro and Ranjay… See the full description on the dataset page: https://huggingface.co/datasets/weikaih/imaginative-perception-token-mvc-answeronly.2026-09-16-odcv-qwen36-0-da-7-answer-only
odcv eval of dougalldeepmind/2026-09-16-qwen36-0-da-7-answer-only (mode=think)
field
value
experiment
odcv eval of dougalldeepmind/2026-09-16-qwen36-0-da-7-answer-only (mode=think)
date_generated
2026-09-16
constitution
none
source_repo
teaching_claude_why_replication @ 0ee0e1248d478f976806b783deb0987c04cbaa5c
models
{"target": "dougalldeepmind/2026-09-16-qwen36-0-da-7-answer-only", "target_revision": "52adc308c378457a94b2eb900802dc424ab5540f", "base":… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-09-16-odcv-qwen36-0-da-7-answer-only.spatial-imaginative-token-pt-answeronly
Spatial Imaginative Token — Path Tracing (Answer-only (label-only; also the answer-only half of mixed training))
Path Tracing (PT) training split for the Spatial Imaginative Token project (11204 samples).
Variant: Answer-only (label-only; also the answer-only half of mixed training).
Used by Spatial-Imaginative-Token:
download with python scripts/download_spatial_datasets.py --task pt.
lvr-data-mvc_answeronly2026-09-16-da-7-answer-only-mix
DA supervision answer; all 752 DA and 9284 identical replay rows
field
value
experiment
DA supervision answer; all 752 DA and 9284 identical replay rows
date_generated
2026-09-16
constitution
constitutions/claude_distilled_09_principles/constitution.md
source_repo
https://github.com/Matthew-Bozoukov/teaching_claude_why_replication.git @ 4648153af4b834b70bd2e5374f639aaad219c83c
models
Tokenizer Qwen/Qwen3.6-27B@6a9e13bd6fc8f0983b9b99948120bc37f49c13e9; replay… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-09-16-da-7-answer-only-mix.Eklav-Reranker-AnswerOnly-Data
Eklav-Reranker-AnswerOnly-Data
Training data for the Eklav paper.
Task: passage reranking (BRIGHT / NevIR benchmarks)
Method: Answer-only (no reasoning of any kind -- the no-CoT floor)
Examples: 381,934 train / 3,857 held-out val
Format: ShareGPT (system + conversations: [{from, value}]), used for LoRA SFT via LLaMA-Factory.
Single-turn ShareGPT conversations. Each row: a query+passage relevance-judgment prompt (human turn) and a bare true/false judgment (gpt turn) -- no hint… See the full description on the dataset page: https://huggingface.co/datasets/AdarshSingh7647/Eklav-Reranker-AnswerOnly-Data.opsd-answeronly-bundle2026-10-03-answeronly-15-mix
answer-only arm: the base blend scaled around a difficult-advice share with every reasoning trace removed
field
value
experiment
answer-only arm: the base blend scaled around a difficult-advice share with every reasoning trace removed — final training mixture (synthetic sources mixed in)
date_generated
20261003
constitution
constitutions/claude_distilled_09_principles/constitution.md
source_repo
git@github.com:Matthew-Bozoukov/teaching_claude_why_replication.git… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-10-03-answeronly-15-mix.answer-only-gp-l-only-10k
Debunk the Myth of SFT Generalization Dataset
This dataset is associated with the paper "Debunk the Myth of SFT Generalization". The paper challenges the prevailing view that supervised fine-tuning (SFT) primarily memorizes training data and fails to generalize, in contrast to reinforcement learning (RL). It demonstrates that SFT can generalize as well as—or better than—RL when trained with appropriate data, achieved through prompt diversity and Chain-of-Thought (CoT) supervision on… See the full description on the dataset page: https://huggingface.co/datasets/Xiaofeng77/answer-only-gp-l-only-10k.2026-10-03-odcv-qwen36-0-answeronly-15
odcv eval of dougalldeepmind/2026-10-03-qwen36-0-answeronly-15 (mode=think)
field
value
experiment
odcv eval of dougalldeepmind/2026-10-03-qwen36-0-answeronly-15 (mode=think)
date_generated
2026-10-03
constitution
none
source_repo
teaching_claude_why_replication @ e38655b136fe0f42f4281d282bab135e777bcdb7
models
{"target": "dougalldeepmind/2026-10-03-qwen36-0-answeronly-15", "target_revision": "057561540d49e26d968537c38e8b462832655031", "base":… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-10-03-odcv-qwen36-0-answeronly-15.imaginative-perception-token-pet-answeronly
Citation
Released with the paper Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models (arXiv:2606.03988):
@misc{bigverdi2026imaginativeperceptiontokensenhance,
title={Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models},
author={Mahtab Bigverdi and Linjie Li and Weikai Huang and Yiming Liu and Jaemin Cho and Jieyu Zhang and Tuhin Kundu and Chris Dangjoo Kim and Zelun Luo and Linda Shapiro and Ranjay… See the full description on the dataset page: https://huggingface.co/datasets/weikaih/imaginative-perception-token-pet-answeronly.2026-10-05-da-15-answer-only-mix
Full-CoT-masked difficult-advice ablation: retain reasoning in context, supervise only the answer, and preserve the October 3 arms replay and corpus; accept the finite-pool token share near 14.82%.
field
value
experiment
Full-CoT-masked difficult-advice ablation: retain reasoning in context, supervise only the answer, and preserve the October 3 arms replay and corpus; accept the finite-pool token share near 14.82%. — final training mixture (synthetic sources mixed in)… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-10-05-da-15-answer-only-mix.lvr-data-path_tracing_answeronly
Spatial Imaginative Token — Path Tracing (Answer-only (label-only; also the answer-only half of mixed training))
Path Tracing (PT) training split for the Spatial Imaginative Token project (11204 samples).
Variant: Answer-only (label-only; also the answer-only half of mixed training).
Used by Spatial-Imaginative-Token:
download with python scripts/download_spatial_datasets.py --task pt.
2026-10-03-da-answeronly-synth
da-answeronly — the difficult-advice corpus with no reasoning
field
value
source
dougalldeepmind/2026-10-02-da-synth @ 305914d58627
rows
1283 (every source row, none dropped)
change
reasoning_content removed from every message; system turn, user turn, assistant reply and metadata byte-identical to the source
answer-only supervised tokens
716,321 (mean 558/row)
share it can fund
14.74% of the published base blend's supervised tokens
declare as
reasoning: none… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-10-03-da-answeronly-synth.lvr-data-pet_answeronly
Citation
Released with the paper Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models (arXiv:2606.03988):
@misc{bigverdi2026imaginativeperceptiontokensenhance,
title={Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models},
author={Mahtab Bigverdi and Linjie Li and Weikai Huang and Yiming Liu and Jaemin Cho and Jieyu Zhang and Tuhin Kundu and Chris Dangjoo Kim and Zelun Luo and Linda Shapiro and Ranjay… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/lvr-data-pet_answeronly.2026-09-01-odcv-answer-only-chunk-only-702-1x65
ODCV-Bench eval of LASR-Callum/qwen3.6-27b-lora-t2-9284-chunk-only-702-answeronly-r64 (mode=think) - the ANSWER-ONLY supervision arm on the principle-scoped (chunk-only) corpus. Its 702 difficult-advice rows train on the VISIBLE ANSWER ONLY: the reasoning trace stays in the token stream as unsupervised context (no truncation) and earns no loss. 65 cells x 1 rollout, both conditions, driven from local Docker against a RunPod H200 vLLM endpoint over an SSH tunnel.
field… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-09-01-odcv-answer-only-chunk-only-702-1x65.genvf-filtered-answer-only-K4-summaries-nextN-prl-trainnq-question-answeronly_addy88_cleanedanswer-only-sokoban
Debunk the Myth of SFT Generalization
This dataset is part of the research presented in the paper Debunk the Myth of SFT Generalization.
The paper challenges the prevailing view that supervised fine-tuning (SFT) memorizes training data and fails to generalize, whereas reinforcement learning (RL) attains broader robustness. Through systematic evaluation on decision-making benchmarks like Sokoban and General Points, the authors demonstrate that introducing prompt diversity and… See the full description on the dataset page: https://huggingface.co/datasets/Xiaofeng77/answer-only-sokoban.genvf-filtered-answer-onlydata_gpt54_only_answer_lossdpo_answer_only_0.05_with_gold_labels_kl_estimationD-sft_gs__structure_types__answer_revision_only__masked_high_lr-sft-datadiverse-answer-only-gp-l-only-10k
General Points Dataset from Debunk the Myth of SFT Generalization
This dataset is part of the research presented in the paper Debunk the Myth of SFT Generalization. It contains data for the General Points decision-making benchmark, which is used to evaluate the generalization capabilities of Supervised Fine-Tuning (SFT) models against Reinforcement Learning (RL) baselines. The paper explores the impact of prompt diversity and Chain-of-Thought (CoT) supervision on SFT's ability to… See the full description on the dataset page: https://huggingface.co/datasets/Xiaofeng77/diverse-answer-only-gp-l-only-10k.dpo_answer_only_with_gold_labels_kl_estimationgenvf-filtered-answer-only-K4-summaries-nextN