datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
dclm-baseline-1.0
DCLM-baseline
DCLM-baseline is a 4T token / 3B document pretraining dataset that achieves strong performance on language model benchmarks.
Below are comparisions of model trained on DCLM-baseline with other models in the 7B regime.
Model
Params
Tokens
Open dataset?
CORE
MMLU
EXTENDED
Open weights, closed datasets
Llama2
7B
2T
✗
49.2
45.8
34.1
DeepSeek
7B
2T
✗
50.7
48.5
35.3
Mistral-0.3
7B
?
✗
57.0
62.7
45.1
QWEN-2
7B
?
✗
57.5
71.9
50.5
Llama3
8B
15T
✗… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations/dclm-baseline-1.0.lm-eval-results-princeton-nlp-Llama-3-Base-8B-SFT-RDPO-private
Dataset Card for Evaluation run of princeton-nlp/Llama-3-Base-8B-SFT-RDPO
Dataset automatically created during the evaluation run of model princeton-nlp/Llama-3-Base-8B-SFT-RDPO
The dataset is composed of 62 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 7 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the… See the full description on the dataset page: https://huggingface.co/datasets/nyu-dice-lab/lm-eval-results-princeton-nlp-Llama-3-Base-8B-SFT-RDPO-private.so100_base_env
SO-100 Base Environment (TsFile)
Apache TsFile version of shreyasgite/so100_base_env.
Overview
A LeRobot teleoperation dataset recorded on an SO-100 arm performing a Lego
pick-and-place task: "Grasp a lego block and put it in the bin." Each episode
captures the synchronized robot joint state and commanded action at every control
step.
Robot: so100 (6-DoF arm: shoulder pan/lift, elbow flex, wrist flex/roll, gripper).
Episodes: 252 (single train split).
Frames: 98… See the full description on the dataset page: https://huggingface.co/datasets/THULab/so100_base_env.BananaMind-Base-Bench-1.1
BananaMind Base Bench 1.1
BananaMind Base Bench 1.1 is an English text-completion benchmark for base causal language models. It contains 350 individually authored examples across seven categories and reports one fixed-scale Overall Elo score.
This is not an instruction-following benchmark. Models receive plain text followed by four possible continuations. The official runner selects the continuation with the highest mean conditional token log-probability. It does not use a chat… See the full description on the dataset page: https://huggingface.co/datasets/BananaMind/BananaMind-Base-Bench-1.1.fixed-n-rb-er-cost-marginrl-qwen3-1.7b-base-math12k-token-mean-fixed-q0p8-run2-rollouts
fixed_n_rb_er_cost_marginrl_Qwen3-1.7B-Base_math12k_token_mean_fixed_q0.8_run2 rollouts
This dataset contains one compressed JSONL shard for every completed training
step. The step and rollout_index columns uniquely locate a rollout within
this training run. Run metadata and per-step row counts are recorded in
rollout_manifest.json.
franka_pnp_big100_base
Franka PnP Big-100 — Base
100 shelf pick-and-place demonstrations with no clutter: one coke can on the floor of a large shelf, a terracotta dish on the desk to the left.
Collected in Isaac Lab (Isaac Lab Arena) with a Franka Panda and a cuRobo-planned scripted expert. One successful episode per scene; the can position is sampled uniformly over the shelf floor (see Sampling). This is one of a pair of datasets: franka_pnp_big100_base (no clutter) and franka_pnp_big100_distract… See the full description on the dataset page: https://huggingface.co/datasets/lithyeon/franka_pnp_big100_base.base-eval-rollouts
base eval rollouts
Eval rollouts (32 samples/problem) for ReasoningRegisters/base (untrained base model).
Split folders named step{k}_{benchmark}[_variant]; JSONL per shard: prompt, generation, correctness.
hh-harmless-base-qwen3-8b-margin-dpo-margin-logsdeepagent
DeepAgent
Hard, Docker-verifiable software-engineering benchmarks from real merged PRs
DeepAgent ships real_pr Harbor hardness packs: live-mined multi-file pull requests, clone@SHA agent images, held-out verifier tests, and Docker dual-truth (solution reward = 1, null reward = 0). Primary product work runs through the deepagent CLI in the GitHub monorepo.
Surface
Ref
Role
HF stable pin
this dataset revision main
Current product on Hub (N=9)
HF automation… See the full description on the dataset page: https://huggingface.co/datasets/BaseIntelligence/deepagent.user_study-preference-personalized_0423_base_filtered
Filtered user study dataset
Source repo: ehejin/user_study-preference-personalized_0423_base
Each row is ONE item review (pre-rating, conversation, post-rating). Submission-level
fields (prolific_pid, demographics, background) are duplicated across rows that share
a submission.
The 25-50 rows here are the FIRST review for each unique pool index, selected the same
way the analysis plot uses — see scripts/plot_vote_shift_3way.py.
Total rows: 50
rollout_output_reward_qwen3_8b_baselm-eval-results-bobofrut-ladybird-base-7B-v8-private
Dataset Card for Evaluation run of bobofrut/ladybird-base-7B-v8
Dataset automatically created during the evaluation run of model bobofrut/ladybird-base-7B-v8
The dataset is composed of 62 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 2 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the latest results.
An… See the full description on the dataset page: https://huggingface.co/datasets/nyu-dice-lab/lm-eval-results-bobofrut-ladybird-base-7B-v8-private.nemiling-knowledge-base
Nemiling Knowledge Base
Nemiling Knowledge Base is the official structured knowledge dataset about Nemiling.
Nemiling is a Russian platform for automating the monetization of Telegram projects through paid subscriptions, paid messages, paid consultations, and donations.
The platform can be used for projects with Russian and international audiences.
The dataset is maintained by the official Nemiling organization and provides structured, machine-readable information about the… See the full description on the dataset page: https://huggingface.co/datasets/nemiling-official/nemiling-knowledge-base.lm-eval-results-princeton-nlp-Llama-3-Base-8B-SFT-IPO-private
Dataset Card for Evaluation run of princeton-nlp/Llama-3-Base-8B-SFT-IPO
Dataset automatically created during the evaluation run of model princeton-nlp/Llama-3-Base-8B-SFT-IPO
The dataset is composed of 62 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 7 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the… See the full description on the dataset page: https://huggingface.co/datasets/nyu-dice-lab/lm-eval-results-princeton-nlp-Llama-3-Base-8B-SFT-IPO-private.fixed-n-rb-cost-aware-marginrl-qwen3-1.7b-base-math12k-token-mean-rerun-rollouts
fixed_n_rb_cost_aware_marginrl_Qwen3-1.7B-Base_math12k_token_mean_rerun rollouts
This dataset contains one compressed JSONL shard for every completed training
step. The step and rollout_index columns uniquely locate a rollout within
this training run. Run metadata and per-step row counts are recorded in
rollout_manifest.json.
bonsai-knowledge-base
bonsAI Knowledge Base
Offline strategy and troubleshooting corpus for bonsAI, a
self-hosted AI assistant plugin for Steam Deck (Decky Loader). This dataset is downloaded at
runtime by the plugin — it is not bundled with the plugin itself, and the plugin (Apache-2.0)
ships no corpus content.
What's in it
117 strategy cards across 13 titles (Baldur's Gate 3, Cyberpunk 2077, Deep Rock Galactic:
Survivor, Fallout 4, Grand Theft Auto: San Andreas — The Definitive… See the full description on the dataset page: https://huggingface.co/datasets/qd313/bonsai-knowledge-base.lm-eval-results-princeton-nlp-Llama-3-Base-8B-SFT-KTO-private
Dataset Card for Evaluation run of princeton-nlp/Llama-3-Base-8B-SFT-KTO
Dataset automatically created during the evaluation run of model princeton-nlp/Llama-3-Base-8B-SFT-KTO
The dataset is composed of 62 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 7 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to… See the full description on the dataset page: https://huggingface.co/datasets/nyu-dice-lab/lm-eval-results-princeton-nlp-Llama-3-Base-8B-SFT-KTO-private.f-cov-offset256-qwen3-1.7b-base-math12k-d72a4f5d-rollouts
f_cov_Qwen3-1.7B-Base_math12k_offset256_token_mean_resume100 rollouts
This dataset contains one compressed JSONL shard for every completed training
step. The step and rollout_index columns uniquely locate a rollout within
this training run. Run metadata and per-step row counts are recorded in
rollout_manifest.json.
fixed-n-rb-er-cost-marginrl-qwen3-1.7b-base-math12k-token-mean-run2-rollouts
fixed_n_rb_er_cost_marginrl_Qwen3-1.7B-Base_math12k_token_mean_run2 rollouts
This dataset contains one compressed JSONL shard for every completed training
step. The step and rollout_index columns uniquely locate a rollout within
this training run. Run metadata and per-step row counts are recorded in
rollout_manifest.json.
2026-09-09-nonmoral-baseline-comparison
Common-protocol ODCV checkpoint comparison
720 scored rollouts; $31.93 estimated total exposure / $300. All owned pods terminated. No new SFT: fresh paired development accepted 13/32 and failed its frozen gate.
Scenario-paired 95% intervals for these three fixed checkpoints. Repeated evaluation passes are not training seeds. Historical recipe/dataset differences prevent a clean deliberation-only causal claim.
Checkpoint
MR
Scenario 95% CI
Submitted
Progress mean /5… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-09-09-nonmoral-baseline-comparison.CYFRAGOVPL__Llama-PLLuM-8B-base-details
Dataset Card for Evaluation run of CYFRAGOVPL/Llama-PLLuM-8B-base
Dataset automatically created during the evaluation run of model CYFRAGOVPL/Llama-PLLuM-8B-base
The dataset is composed of 38 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the latest results.… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/CYFRAGOVPL__Llama-PLLuM-8B-base-details.lm-eval-results-princeton-nlp-Llama-3-Base-8B-SFT-private
Dataset Card for Evaluation run of princeton-nlp/Llama-3-Base-8B-SFT
Dataset automatically created during the evaluation run of model princeton-nlp/Llama-3-Base-8B-SFT
The dataset is composed of 62 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 2 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the latest… See the full description on the dataset page: https://huggingface.co/datasets/nyu-dice-lab/lm-eval-results-princeton-nlp-Llama-3-Base-8B-SFT-private.dapo-math-17k-difficulty-qwen3-1.7b-base-k16
DAPO-Math-17k difficulty under Qwen3-1.7B-Base (K=16)
For each of the 17,398 problems in the DAPO-Math-17k train set, how many of
K=16 samples from the untrained base model are correct.
The headline: 57.27% of problems are solved 0 out of 16 times, and not one
problem is solved 16 out of 16. Difficulty here is entirely one-sided.
Why count per problem instead of reporting mean accuracy
In group-relative RL (GRPO and its relatives), a prompt group whose K responses… See the full description on the dataset page: https://huggingface.co/datasets/RyanYr/dapo-math-17k-difficulty-qwen3-1.7b-base-k16.fixed-n-rb-offset256-qwen3-1.7b-base-math12k-754f8ca2-rollouts
fixed_n_rb_offset_cost_aware_marginrl_Qwen3-1.7B-Base_math12k_offset256_token_mean rollouts
This dataset contains one compressed JSONL shard for every completed training
step. The step and rollout_index columns uniquely locate a rollout within
this training run. Run metadata and per-step row counts are recorded in
rollout_manifest.json.
f-cov-offset1024-qwen3-1.7b-base-math12k-c50b0bba-rollouts
f_cov_Qwen3-1.7B-Base_math12k_offset1024_token_mean rollouts
This dataset contains one compressed JSONL shard for every completed training
step. The step and rollout_index columns uniquely locate a rollout within
this training run. Run metadata and per-step row counts are recorded in
rollout_manifest.json.
Baseline_featbeachf-cov-offset2048-qwen3-1.7b-base-math12k-4a578b3f-rollouts
f_cov_Qwen3-1.7B-Base_math12k_offset2048_token_mean rollouts
This dataset contains one compressed JSONL shard for every completed training
step. The step and rollout_index columns uniquely locate a rollout within
this training run. Run metadata and per-step row counts are recorded in
rollout_manifest.json.
bert-base-multilingual-cased-toksuite-detokenizedTraining data of the model detokenized in the exact order seen by the model.
The training data is partitioned into 8 chunks (chunk-0 through chunk-7), based on the GPU rank that generated the data. Each chunk contains detokenized text files in JSON Lines format (.jsonl).
ibm-granite__granite-3.0-2b-base-details
Dataset Card for Evaluation run of ibm-granite/granite-3.0-2b-base
Dataset automatically created during the evaluation run of model ibm-granite/granite-3.0-2b-base
The dataset is composed of 38 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 2 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the latest… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/ibm-granite__granite-3.0-2b-base-details.mistralai__Mistral-Small-24B-Base-2501-details
Dataset Card for Evaluation run of mistralai/Mistral-Small-24B-Base-2501
Dataset automatically created during the evaluation run of model mistralai/Mistral-Small-24B-Base-2501
The dataset is composed of 38 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 2 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/mistralai__Mistral-Small-24B-Base-2501-details.