datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Proxy3D-SpaceSpan-318K
SpaceSpan Dataset
SpaceSpan is a large-scale dataset curated for the training and evaluation of 3D vision-language models (VLMs), specifically introduced in the paper Proxy3D: Efficient 3D Representations for Vision-Language Models via Semantic Clustering and Alignment.
Project Page | GitHub Repository
Dataset Description
The SpaceSpan dataset is designed to help VLMs develop spatial intelligence through 3D proxy representations. It incorporates heterogeneous visual… See the full description on the dataset page: https://huggingface.co/datasets/Spacewanderer8263/Proxy3D-SpaceSpan-318K.proxy-listproxy-logs-ShareGPTAll the proxy logs I could find (lmk if there are more), converted to ShareGPT so it's all formatted the same way in one dataset. I didn't .strip() any of the turns, I only used ftfy.fix_text() on them, and skipped any empty turns.
The original sample response is moved to be the final turn of conversations. If the last turn in the original sample prompt was a model turn, it is assumed that it was for doing prefill. So I moved this turn to response_prefill.
sample["response_prefill"] and… See the full description on the dataset page: https://huggingface.co/datasets/PJMixers-Dev/proxy-logs-ShareGPT.greater-london-tx-proxy-rat-path-gain
Greater London Per-Transmitter-Proxy, Per-RAT Simulated Path-Gain Dataset
Short display name: Greater London Tx-Proxy × RAT Path-GainChinese name: 大伦敦逐发射代理、逐 RAT 模拟路径增益数据集Release: v9 final release (COMPLETE)
City-scale propagation, one transmitter proxy at a time.
This release turns Greater London into a queryable radio-propagation dataset:
449,437,201,731 simulated path-gain relations connect 22,678 computed
transmitter-proxy hypotheses with 171,549,960 receiver faces across… See the full description on the dataset page: https://huggingface.co/datasets/EEzim/greater-london-tx-proxy-rat-path-gain.ProxyCoT-HotpotQAThis is the HotpotQA data that we used in our ProxyCoT project (https://aclanthology.org/2026.acl-long.1917/), and it is based on long-context reasoning (32K-128K tokens).
HotpotQA here is a new version originally from https://aclanthology.org/2026.acl-long.1917/ with extended contexts.
For more details on the context extension, refer to the ProxyCoT paper.
To use our dataset, please follow the code below.
train_samples = load_dataset("oaimli/proxycot-hotpotqa", split="train")
dev_samples =… See the full description on the dataset page: https://huggingface.co/datasets/oaimli/ProxyCoT-HotpotQA.ProxyCoT-SciTrekThis is the SciTrek data that we used in the ProxyCoT project, and it is based on long-context reasoning (32K-128K tokens).
SciTrek is originally from https://arxiv.org/abs/2509.21028.
To use our dataset, please follow the code below.
train_samples = load_dataset("oaimli/proxycot-scitrek", split="train")
dev_samples = load_dataset("oaimli/proxycot-scitrek", split="val")
test_samples = load_dataset("oaimli/proxycot-scitrek", split="test")
for sample in train_samples:
question =… See the full description on the dataset page: https://huggingface.co/datasets/oaimli/ProxyCoT-SciTrek.openve_cfg_3arm_proxyhuper-clean100-proxyphones
huper-clean100-proxyphones
LibriSpeech train-clean-100 audio paired with HuPER-style proxy ARPAbet phone labels (machine-generated / proxy, not human verified).This corresponds to the 100h train-clean-100 split (28,539 utterances).Note: HuggingFace Dataset Viewer is not supported because the data is provided as tar+zstd shards. Follow the instructions below to download and extract locally.
What’s inside
The data is stored as 5 shards under blobs/:… See the full description on the dataset page: https://huggingface.co/datasets/huper29/huper-clean100-proxyphones.combust-labs_pi-mono-http-proxyproxy-logs-ReRolls-Minos
non-refusal responses: 845,186
refusal responses: 38,285
https://gist.github.com/xzuyn/1d7f43db2750060a18a304eb84b396db
Use a training prompt formatter like this: https://github.com/xzuyn/axolotl/blob/latest-formatters/src/axolotl/prompt_strategies/customllama3-regex-last-only-prefill-reroll.py
free-proxy-observations
Historical Litport free proxy observations
This is a fixed historical snapshot observed at 2026-09-15T19:36:14.715Z, containing 1519 publicly listed proxy observations. It is not a live feed and carries no freshness badge.
Data
all.json and all.csv contain the same public schema fields.
schema.json records field names, types, and nullability.
provenance.json records the source commit and SHA-256 checksums.
Source: litportnet/free-proxy-list commit… See the full description on the dataset page: https://huggingface.co/datasets/litportnet/free-proxy-observations.proxy-logs-ReRollsDuplicate prompts combined into a single sample, with all responses in a list of dicts. I've also included some info like token count, and slop (though my slop list could use improvement).
Use a training prompt formatter like this: https://github.com/xzuyn/axolotl/blob/84aec029dfa9eb9670b8a51d432a279be6c85871/src/axolotl/prompt_strategies/customllama3-regex-last-only-prefill-reroll.py
Dataset creation script: https://gist.github.com/xzuyn/aa1f30b7394d2997766bef82edb67227
herculaneum-legibility-proxy-labels
Herculaneum legibility proxy — training labels (proxy_v4 / proxy_s2s3_v1)
This is the label set and training recipe behind the proxy_v4 /
proxy_s2s3_v1 model card:
https://huggingface.co/LimeGS/herculaneum-legibility-proxy.
Read that first — it documents the model, the preprocessing, the
thresholds, and the validation results. This dataset only covers how the
model was trained: the human labeling protocol, the exact windows it
produced, and a script to turn them back into pixels… See the full description on the dataset page: https://huggingface.co/datasets/LimeGS/herculaneum-legibility-proxy-labels.proxy-ip-pricing-cn
Proxy IP Pricing (China Market) 2026
An open dataset of proxy IP pricing, protocol support, coverage and official registration links for 18 providers serving the Chinese market. Compiled monthly from provider-published price sheets.
Maintained by 全网低价IP / socks5ip — a comparison platform aggregating 20+ proxy IP providers.
Why this dataset exists
Proxy IP pricing is scattered across dozens of provider sites, quoted in different units (per day / per week / per… See the full description on the dataset page: https://huggingface.co/datasets/socks5ip/proxy-ip-pricing-cn.TCGA-OT-ProxyModel-Outputs
TCGA-OT Proxy Model Outputs (UNI, CONCH)
This dataset contains the raw output dictionaries from two proxy models (conch and uni) trained on TCGA-OT data. The outputs are stored as Python pickle files.
Each file contains a dictionary with a single key, "split0". The value for this key is another dictionary where keys are TCGA slide_ids and values are the model's predictions for that slide.
Dataset Structure
The dataset consists of two files:… See the full description on the dataset page: https://huggingface.co/datasets/ownt/TCGA-OT-ProxyModel-Outputs.Proxy3D-annotations
SpaceSpan Dataset
SpaceSpan is a large-scale dataset curated for aligning 3D proxy representations with Vision-Language Models (VLMs), introduced in the paper Proxy3D: Efficient 3D Representations for Vision-Language Models via Semantic Clustering and Alignment.
The dataset incorporates heterogeneous visual information into a unified format to support multi-stage training for developing spatial intelligence. It enables models to progress from simple image-text alignment to complex… See the full description on the dataset page: https://huggingface.co/datasets/Spacewanderer8263/Proxy3D-annotations.proxyquotes_library
The Proxy Quotes (pxyq) library
includes a fuction for calling the cell value with respect to a column and row of the csv dataset table. It calls for proxy stoploss distance, lotsize, and margins with leverages covering a betsize of 1 cash, commissions, swaps, spread, and more. It only have one simple function call, and that is pxyq.column('ASSET').
step 1: make sure you have pxyq.py in your directory. No need pip installations.
step 2: make an import pxyq is written on top of… See the full description on the dataset page: https://huggingface.co/datasets/algorembrant/proxyquotes_library.ecup-2026-matching-eval-proxy-v2
E-CUP 2026 retrospective evaluation proxy v2
Private team dataset for reproducing the checksum-bound retrospective contest transfer diagnostic.
Primary panel
file: soft_sample_pairs.parquet
immutable revision: 183601763fd7f4d1695325b315ef3c7cc98e67c1
immutable download: https://huggingface.co/datasets/mariklolik/ecup-2026-matching-eval-proxy-v2/resolve/183601763fd7f4d1695325b315ef3c7cc98e67c1/soft_sample_pairs.parquet
rows: 250000
SHA-256:… See the full description on the dataset page: https://huggingface.co/datasets/mariklolik/ecup-2026-matching-eval-proxy-v2.proxy-trajfullstack-claim2-proxyproxy-ip-pricing-cn-2026
Proxy IP Pricing (China Market) 2026
An open dataset of proxy IP pricing, protocol support, coverage and official registration links for 18 providers serving the Chinese market. Compiled monthly from provider-published price sheets.
Maintained by 全网低价IP / socks5ip — a comparison platform aggregating 20+ proxy IP providers.
Why this dataset exists
Proxy IP pricing is scattered across dozens of provider sites, quoted in different units (per day / per week / per… See the full description on the dataset page: https://huggingface.co/datasets/socks5ip/proxy-ip-pricing-cn-2026.g-vendi-base-proxy
Use a base-model proxy for G-Vendi, and score it within format
Companion data for the FinePhrase Synthetic Data Playbook (Niklaus et al., 2026),
on the 83-cell grid from its Figure 22. G-Vendi (Jung et al., 2025) is a
gradient-space diversity metric that the playbook uses as one of several
predictors of downstream performance.
The heatmap correlates each predictor with each benchmark across the grid. The
top three blocks are the playbook's other predictors (DCLM, Edu, and the… See the full description on the dataset page: https://huggingface.co/datasets/ratishsp/g-vendi-base-proxy.OR-synthetic-proxy
OR Scheduling Synthetic Proxy Dataset
This synthetic dataset accompanies the paper:
Decision-Focused Learning for Operating Room Scheduling Under Uncertain
Surgery Durations with Two-Stage Stochastic OptimizationAli Elaswad, Rodrigo Carrasco, Nourhan Sakr — ECML-PKDD 2026
Description
A synthetic proxy dataset generated from anonymized aggregate statistics
of a real neurosurgical dataset from the Instituto de Neurocirugia Dr.
Raul Asenjo, Santiago, Chile.… See the full description on the dataset page: https://huggingface.co/datasets/alyelaswad/OR-synthetic-proxy.il_gym_stateThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.1",
"robot_type": null,
"total_episodes": 30,
"total_frames": 1106,
"total_tasks": 1,
"total_videos": 0,
"total_chunks": 1,
"chunks_size": 1000,
"fps": 10,
"splits": {
"train": "0:30"},
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/ProxyCausal/il_gym_state.il_gym1This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.1",
"robot_type": null,
"total_episodes": 50,
"total_frames": 1443,
"total_tasks": 1,
"total_videos": 100,
"total_chunks": 1,
"chunks_size": 1000,
"fps": 10,
"splits": {
"train": "0:50"},
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/ProxyCausal/il_gym1.il_gym0This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.1",
"robot_type": null,
"total_episodes": 30,
"total_frames": 1610,
"total_tasks": 1,
"total_videos": 60,
"total_chunks": 1,
"chunks_size": 1000,
"fps": 10,
"splits": {
"train": "0:30"},
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/ProxyCausal/il_gym0.Proxy-GSil_gym_state_overfitThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.1",
"robot_type": null,
"total_episodes": 1,
"total_frames": 54,
"total_tasks": 1,
"total_videos": 0,
"total_chunks": 1,
"chunks_size": 1000,
"fps": 10,
"splits": {
"train": "0:1"},
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/ProxyCausal/il_gym_state_overfit.pick_and_place_SO101This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "so101_follower",
"total_episodes": 12,
"total_frames": 5424,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 500,
"fps": 30,
"splits": {
"train": "0:12"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/ProxyCausal/pick_and_place_SO101.eval_so101This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "so101_follower",
"total_episodes": 3,
"total_frames": 2460,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 500,
"fps": 30,
"splits": {
"train": "0:3"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/ProxyCausal/eval_so101.