datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
aetheris-experiencesAetherSearch_Eval_1400
🔭 AetherSearch Eval-1400
One frozen benchmark for training-time evaluation and final checkpoint assessment
🏠 Project ·
🎓 SFT Data ·
🤖 SFT Model ·
⚖️ DPO Data ·
🧠 DPO Model
Dataset overview
AetherSearch Eval-1400 is a frozen, 1,400-question evaluation suite for
agentic search. It combines seven official held-out QA sources and isolates
their questions from the audited AetherSearch SFT, DPO, and RL training inputs.
This… See the full description on the dataset page: https://huggingface.co/datasets/muradil211/AetherSearch_Eval_1400.AetherCode
AetherCode: Evaluating LLMs' Ability to Win In Premier Programming Competitions
Introduction
Competitive programming has emerged as a critical benchmark for evaluating the reasoning and coding capabilities of Large Language Models (LLMs). Despite impressive progress on existing benchmarks, we argue that current evaluations overstate model proficiency, masking a substantial gap between LLMs and elite human programmers. This gap arises… See the full description on the dataset page: https://huggingface.co/datasets/m-a-p/AetherCode.aether-sft-v1-sources
aether-sft-v1-sources
Top-tier generalist SFT instruction-tuning sources for AETHER training. Aggregates the SOTA datasets: Tulu-3 SFT mixture (Allen AI), OpenHermes-2.5 (Teknium), NuminaMath-CoT/1.5 (AI-MO, math reasoning), WildChat-1M (real GPT-4 conversations), Dolphin + Dolphin-R1 (reasoning), Tulu-3 personas (math/instr). Multi-skill: instruction-following, math reasoning, coding, dialogue, multilingual.
Disclaimer (Responsible Disclosure)
This bundle aggregates… See the full description on the dataset page: https://huggingface.co/datasets/jescy525/aether-sft-v1-sources.Aether-V1.9
Aether Dataset
Creator: SteelSkull
About Aether: The Aether dataset.
Rebuilt script from v1.8.5 to v1.9.
Version v1.9
Due to an error in the codebase the 'system' and 'tools' records were not being carried over to the final dataframe, it has been fixed
Recommendation from a discord user (#nguyenzzz [they also found the error above]) was to add an 'origins' records for where the dataset was being pulled… See the full description on the dataset page: https://huggingface.co/datasets/TheSkullery/Aether-V1.9.AetherSearch_SFT
AetherSearch Search-SFT 2600
Dataset Overview
This release contains 2,600 validated full agent trajectories for the
Qwen2.5-3B AetherSearch cold start, including retrieval and zero-search
direct-answer trajectories.
The DeepSeek teacher ran with thinking disabled and with no tools registered.
Teacher answers were accepted only when their normalized minimal answer matched
an isolated reference alias. The exported final <think> was canonicalized to
the public… See the full description on the dataset page: https://huggingface.co/datasets/muradil211/AetherSearch_SFT.akie-pretrain-corpus
AKIE Pretrain Corpus
Corpus de pré-treino para a família de modelos AKIE, organizado em 4
eixos com proporções fixas, tokenizado com o
AkieTokenizer
(SentencePiece BPE, 32k vocabulário).
Composição
Eixo
Proporção
Tokens
Código
40%
~2,40B
Instruções
20%
~1,20B
Diálogo
25%
~1,50B
Raciocínio
15%
~0,90B
Total
100%
~6,00B
Fontes: código-fonte de repositórios públicos (várias linguagens),
diálogos e instruções em português (traduções e coleções… See the full description on the dataset page: https://huggingface.co/datasets/AETHER-LAB/akie-pretrain-corpus.cpp_cwe_GRPO
cpp_cwe_GRPO
VeRL/GRPO-ready C++ security-coding dataset produced from every harness-passing stage-6 rewrite in the simple_gen pipeline.
Files
File
Rows
Description
cpp_cwe_GRPO.parquet
8,039
Full corrected dataset
cpp_cwe_GRPO_train.parquet
7,236
Deterministic 90% training split
cpp_cwe_GRPO_val.parquet
803
Deterministic 10% validation split
The split uses seed 20260929. Each record has a unique function name, and the two splits are disjoint.… See the full description on the dataset page: https://huggingface.co/datasets/AetherPrior/cpp_cwe_GRPO.3DRAG-Bench
3DRAG-Bench
This dataset contains 100 curated 3D object assets for 3DRAG 3D editing experiments.
Each object is stored as a GLB mesh together with a cleaned editing specification.
Dataset Structure
.
+-- README.md
+-- LICENSE
+-- .gitattributes
+-- metadata.csv
+-- name_mapping.csv
`-- assets/
`-- <asset_name>/
+-- model.glb
`-- dataset_input_clean.json
Files
assets/<asset_name>/model.glb: GLB asset file.… See the full description on the dataset page: https://huggingface.co/datasets/AeTherRaIn/3DRAG-Bench.Aether-V1.5
Aether Dataset
Creator: SteelSkull
Community Organization: ConvexAI
Discord: Join us on Discord
About Aether: The Aether dataset.
rebuilt script, new dataset
from 1.2.2 to 1.5, changed datasets, added two.
version v1.5 is a rework of the human -> gpt conversations and added system and tool columns
Source Datasets:… See the full description on the dataset page: https://huggingface.co/datasets/TheSkullery/Aether-V1.5.c_cwe_GRPO
c_cwe_GRPO
VeRL/GRPO-ready C security coding dataset generated by the simple_gen pipeline.
Each row is a harness-validated task with pytest security/functionality tests and oracle candidate_c. Most CWEs are post stage-6 rubric rewrite (6_rewrites.jsonl); CWE-476 is still from stage-5 guidelines. Stage-7 guideline resampling has not been applied yet, so high_level_guidelines / implementational may be empty on rewritten rows.
Files
File
Rows
Description… See the full description on the dataset page: https://huggingface.co/datasets/AetherPrior/c_cwe_GRPO.AetherCode-v1
Dataset Description
Abstract
The "AetherCode" dataset is designed to fine-tune models on coding tasks across various programming languages, incorporating complex real-world coding scenarios. It aims to push the boundaries of AI in code generation and software development.
How to Load This Dataset
from datasets import load_dataset
dataset = load_dataset("thesven/AetherCode-v1", split="5star")
Languages
The dataset includes coding problems in… See the full description on the dataset page: https://huggingface.co/datasets/thesven/AetherCode-v1.RL_seccode_gen
C++ update
The C++ portion was replaced with the 7,236-row training split from AetherPrior/cpp_cwe_GRPO, built from the harness-passing stage-6 rewritten C++ records. Non-C++ rows were retained unchanged.
C++ guideline refresh
The existing 7,236 C++ training rows retain their exact function-ID set. Revised high-level and implementation guidelines were overlaid for 2,256 matching records from the C++ rewritten-guidelines artifacts; 1,072 records received changed… See the full description on the dataset page: https://huggingface.co/datasets/AetherPrior/RL_seccode_gen.AetherSearch_DPO
🔭 AetherSearch DPO
Preference pairs for reasoning, retrieval, and evidence-grounded answers
🏠 Project ·
🧠 DPO Model ·
🧪 Training Code ·
🎓 SFT Data ·
🤖 SFT Model
Dataset overview
AetherSearch DPO contains 2,126 preference pairs for training an agentic-search
policy after supervised fine-tuning. Every row provides one shared prompt, a
preferred assistant continuation, and a non-preferred continuation.
This repository is… See the full description on the dataset page: https://huggingface.co/datasets/muradil211/AetherSearch_DPO.py_cwe_GRPO
py_cwe_GRPO
Python CWE GRPO training dataset with rubric-aware authoring guidelines, built from the post–step-6 pipeline (6_rewritten.jsonl / 6_rewritten_guidelines.jsonl).
Size
7,977 rows (17 CWEs) — the post-rewrite oracle set, not the older 10k HF aggregate.
How guidelines were produced
Reused rubric-aware guidelines from a prior generation when (cwe, function_name, prompt) matched and generated_code was identical (~5.4k rows).
Regenerated… See the full description on the dataset page: https://huggingface.co/datasets/AetherPrior/py_cwe_GRPO.python_cwe_GRPO
Overview
VeRL/GRPO-ready RL dataset for Python CWE tasks built from simple_gen pipeline outputs.
How it was built
Generated by:
secure_reasoning/security-test-case/simple_gen/py/5_gather_rl_datasets_per_cwe.py --lang python --output-suffix cwe_grpo
Input source:
secure_reasoning/security-test-case/simple_gen/data/pipeline_runs/python/cwe-*/6_rewritten_guidelines.jsonl
Files
python_cwe_grpo.parquet (all rows)
python_cwe_grpo_train.parquet (90%)… See the full description on the dataset page: https://huggingface.co/datasets/AetherPrior/python_cwe_GRPO.SecCodePLTaether-cyber-sft
enosislabs/aether-cyber-sft
Version: surface-clean-20260620
Generated UTC: 2026-06-20T03:36:32.011333+00:00
Source path: artifacts/aether-cyber-sft-surface-candidate.jsonl
Git commit: 6fadc0ba97b4c49a7b644e70afacddc8ba98029e
Curated Aether PRISM SFT dataset for authorized cybersecurity training.
Each source shard follows 5-row review discipline before publish.
Records
Total examples: 1794
Domains
vulnerability_research: 757
red_team_ops: 334… See the full description on the dataset page: https://huggingface.co/datasets/enosislabs/aether-cyber-sft.Aether-minikotodama-3b-corpus-open
kotodama-3b pretraining corpus — open subset
The deduplicated, cleaned text of the openly licensed sources used to pretrain the kotodama 3B
(aethera-gp/kotodama-3b-base-final). 19 of the model's 32 sources; the other 13 (non-commercial,
unlicensed, gated, or flagged sources) are not redistributed here.
Layout: data/<source>/part-NNNN.jsonl.zst — zstd JSON Lines, one document per line.
Processing (github.com/LuxiaSL/kotodama, curation): Unicode normalisation, length filtering… See the full description on the dataset page: https://huggingface.co/datasets/aethera-gp/kotodama-3b-corpus-open.CWE-Code_Vulnerability_Security_DPOAether-Lite-PurHyDe
Aether Lite Dataset
Creator: SteelSkull
About Aether-Lite-PurHyDe: The Aether-Lite dataset is designed to balance creative writing, Slop, and intelligence.
Whats New?:
Aether-Lite-PurHyDe
This dataset is basically a HEAVILY cleaned and filtered version of Aether-lite. ONLY english, ANY and all AI-isms (claud, gpt, gemma) were stripped out and agressive fussy dedupe was applied
Fuzzy deduplication was set to a 90%… See the full description on the dataset page: https://huggingface.co/datasets/TheSkullery/Aether-Lite-PurHyDe.aetherx-port-congestion-metrics
Aether-X Global Port Congestion Snapshot
Point-in-time snapshot of the Aether-X Port Congestion Oracle — predictive
congestion, ETA delay and freight-volatility signals for 15 of the world's
largest ports.
This static CSV is a frozen snapshot for research, backtesting and
dashboards. The live, continuously-updated signal is available through the
REST API and the Python SDK.
Files
port_metrics.csv — one row per port.
Schema
Column
Type… See the full description on the dataset page: https://huggingface.co/datasets/Aether-x/aetherx-port-congestion-metrics.atc-tts-mos-ratingsaether-evals-runsDaemontatox__AetherSett-details
Dataset Card for Evaluation run of Daemontatox/AetherSett
Dataset automatically created during the evaluation run of model Daemontatox/AetherSett
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 results.
An additional… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/Daemontatox__AetherSett-details.aether-elite-mimo
Aether Elite Distillation — mimo
LeeWorld Aether teacher-distillation ELITE set (funnel-gated). Teacher: mimo.
Rows: 3000 — survived the 13-stage quality gauntlet (Nemotron-CC + Instella/Dolma categories).
Method: FuseChat-3.0 IMF — responses sampled from the teacher, scrubbed to our schema, quality-funneled.
Provenance: teacher is Tier-A open-weights (outputs clean). Proprietary — LeeWorld.
Aether-Lite-v1.8.1
Aether Lite Dataset
Creator: SteelSkull
About Aether-Lite-V1.8.1: The Aether-Lite dataset is designed to balance creative writing, Slop, and intelligence.
Whats New?:
1.8 --> 1.8.1: had to strip out the 'token_distribution' column as it was causing issues
New functions added to the script include dataset use percentage (will only use a percentage of the dataset supplied), dataset shuffling, and a new fuzzy deduplication… See the full description on the dataset page: https://huggingface.co/datasets/TheSkullery/Aether-Lite-v1.8.1.Aether-Lite-v1.8
Aether Lite Dataset
Creator: SteelSkull
About Aether-Lite-V1.8: The Aether-Lite dataset is designed to balance creative writing, Slop, and intelligence.
New functions added to the script include dataset use percentage (will only use a percentage of the dataset supplied), dataset shuffling, and a new fuzzy deduplication method on the overall dataset.
The new fuzzy deduplication method was set to a 95% threshold, and I had to… See the full description on the dataset page: https://huggingface.co/datasets/TheSkullery/Aether-Lite-v1.8.Daemontatox__AetherDrake-SFT-details
Dataset Card for Evaluation run of Daemontatox/AetherDrake-SFT
Dataset automatically created during the evaluation run of model Daemontatox/AetherDrake-SFT
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 results.
An… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/Daemontatox__AetherDrake-SFT-details.