datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
nvidia_open_reasoning_balanced_100k
nvidia_open_reasoning_balanced_100k
A domain-balanced 100k reasoning SFT dataset built from three NVIDIA Open Reasoning
datasets: 33,333 examples each for math, code, and science (99,999 total).
Each example is a single-turn conversation with a full reasoning trace:
conversations: [
{"from": "human", "value": "<problem>"},
{"from": "gpt", "value": "<think>\n<reasoning trace>\n</think><final solution>"}
]
Columns
column
description
conversations… See the full description on the dataset page: https://huggingface.co/datasets/bhheo/nvidia_open_reasoning_balanced_100k.route_balance_model_estimator
RouteBalance Model-Estimator Dataset
Prompts broadcast across every model in a serving pool, with per-model responses and
reference-grounded judge scores. Each record gives directly comparable outputs for the same
input, which is what makes the corpus usable for routing research — learning which model in
a heterogeneous pool will answer a given prompt well — rather than only for single-model
evaluation.
Two model families are provided. They share one prompt taxonomy and one… See the full description on the dataset page: https://huggingface.co/datasets/asdwb/route_balance_model_estimator.tinystories-instruct-balanced
Dataset Card for TinyStories Instruct - Balanced
Dataset Summary
TinyStories Instruct - Balanced is a curated, instruction-tuning dataset derived from roneneldan/TinyStoriesInstruct. It contains short story generation examples with balanced happy/sad endings (50-50 split), making it ideal for fine-tuning language models to follow instructions and generate contextually appropriate narratives.
The dataset was created to address the original TinyStoriesInstruct's imbalance… See the full description on the dataset page: https://huggingface.co/datasets/0rn0/tinystories-instruct-balanced.prism_trial_3_balanced
PRISM Trial 3: Fixed Balanced Cohorts
This is the preregistration-ready companion to Alberto1231/prism_trial_3.
Every conversation is dated 2023 or later; the observed range is November
22 through December 22, 2023. Every target is the genuine next human turn after
the assistant response selected by that participant.
Evaluation versus analysis
Use the full configuration for model evaluation. It contains the same 456
unique held-out respondents as PRISM Trial 3, so… See the full description on the dataset page: https://huggingface.co/datasets/Alberto1231/prism_trial_3_balanced.qwen3.8-max-glm5.2-kimi-k3-sft-balanced
Multi-Teacher SFT Balanced Dataset (57,937 Traces)
Quality-filtered, deduplicated, multi-teacher SFT corpus packaged from r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation (subset: sft_balanced).
Dataset Overview
Source Dataset: r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation
Subset: sft_balanced
Total Traces: 57,937 (under the 100k cap)
Standardized Column: The conversation turns are strictly standardized under the messages column (resolved and mapped from… See the full description on the dataset page: https://huggingface.co/datasets/bunnycore/qwen3.8-max-glm5.2-kimi-k3-sft-balanced.2026-08-08-table2-9000-synthdoc-1000-trait-balanced-len-8000-train-mixture
Table-2 (9,000) + synthdoc difficult-advice (1,000, trait-balanced), all rows <= 8,000 tokens
10,000-example SFT mixture for Qwen3.6-27B. Train on mixture_think.jsonl — every
assistant turn carries a think block, which the trainer's preserve-thinking gate requires.
field
value
experiment
90/10-by-examples SFT mixture: 9,000 spec-filtered Table-2 instruction rows + 1,000 difficult-advice documents drawn evenly across all 9 constitution traits
date_generated… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-08-08-table2-9000-synthdoc-1000-trait-balanced-len-8000-train-mixture.2026-08-08-table2-9000-synthdoc-1000-trait-balanced-train-mixture
Table-2 (9,000) + synthdoc difficult-advice (1,000, trait-balanced)
10,000-example SFT mixture for Qwen3.6-27B. Train on mixture_think.jsonl — every
assistant turn carries a think block, which the trainer's preserve-thinking gate requires.
field
value
experiment
90/10-by-examples SFT mixture: 9,000 spec-filtered Table-2 instruction rows + 1,000 difficult-advice documents drawn evenly across all 9 constitution traits
date_generated
2026-08-08
constitution… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-08-08-table2-9000-synthdoc-1000-trait-balanced-train-mixture.The-Balanced-Budget-And-Emergency-Deficit-Control-Act-of-1985
Dataset Description
The Balanced Budget and Emergency Deficit Control Act of 1985 Question-Answer Dataset is an English-language instructional dataset derived from the statutory provisions of the Balanced Budget and Emergency Deficit Control Act of 1985.
The Act was enacted as Title II of Public Law 99-177 on December 12, 1985, and is commonly known as the Gramm-Rudman-Hollings Act. Its provisions established federal budget-enforcement mechanisms intended to control deficits… See the full description on the dataset page: https://huggingface.co/datasets/leeroy-jankins/The-Balanced-Budget-And-Emergency-Deficit-Control-Act-of-1985.gaokao-sft-chinese-balanced
Gaokao SFT Chinese Balanced
This dataset is a cleaned SFT-style Chinese exam dataset prepared from multiple public Hugging Face sources.
Composition
Total samples: 1895
Train samples: 1853
Validation samples: 42
Fields
Each row contains:
id
lang
subject
source
instruction
input
output
messages
Cleaning Notes
Ordinary Markdown markers were removed.
Non-essential LaTeX commands were simplified into plain readable text.
Math expressions were… See the full description on the dataset page: https://huggingface.co/datasets/callofthenight1/gaokao-sft-chinese-balanced.multilingual-instruct-balancedThis repository is a collection of English, Korean, Chinese, and Japanese datasets collected by the HuggingFace Hub and transformed into a unified format. It consists of either native or synthetic data.
Some data is not clearly copyrighted or only allows non-commercial use.
Preprocessing: I removed data with too few answer tokens or more than 8192 tokens, and removed synthetic data with repetitions.
Balancing: I randomly sampled a subset of the data with different weights for each language and… See the full description on the dataset page: https://huggingface.co/datasets/werty1248/multilingual-instruct-balanced.Code-Vulnerability-Balanced
Code Vulnerability Balanced — CWE-Enriched Conversation Dataset
📌 Overview
This dataset is a balanced and shuffled version of
ChamaraVishwajithRajapaksha/Code-Vulnerability-FineTune,
which itself was derived from the original
ChamaraVishwajithRajapaksha/Code_Vulnerability_Dataset
(330k rows, sourced from DiverseVul + MITRE CWE enrichment).
The original fine-tuning dataset was imbalanced — the number of Vulnerable and Safe
samples were not equal — and the samples were not… See the full description on the dataset page: https://huggingface.co/datasets/ChamaraVishwajithRajapaksha/Code-Vulnerability-Balanced.cmmc-training-balanced
CMMC Training Dataset - Balanced Variant
Dataset Description
This is the Balanced variant of the CMMC (Cybersecurity Maturity Model Certification) training dataset, containing 2,790 high-quality training examples with balanced coverage across all 17 CMMC domains.
Dataset Characteristics
Total Examples: 2,790 (2,232 train / 558 validation)
Source Documents: 71 NIST publications
CMMC Levels Covered: Level 1, Level 2, Level 3
CMMC Domains: All 17 domains (evenly… See the full description on the dataset page: https://huggingface.co/datasets/ethanolivertroy/cmmc-training-balanced.toricblm-dataset-state-toricblm-structure-priority-balanced-3day-20260709t185034z-epoch-001
ToricBLM dataset state: toricblm-structure-priority-balanced-3day-20260709T185034Z epoch 001
This dataset repo records the exact local training-data state visible to the dynamic epoch launcher.
It intentionally stores manifests and audit records rather than duplicating large Parquet shards.
Special checkpoint: toricblm-structure-priority-balanced-3day-20260709T185034Z_epoch_001_special_structure_current_step_002000.pt
Checkpoint repo: AmelieSchreiber/ToricGT_160M_FoT
Curriculum… See the full description on the dataset page: https://huggingface.co/datasets/AmelieSchreiber/toricblm-dataset-state-toricblm-structure-priority-balanced-3day-20260709t185034z-epoch-001.toricblm-dataset-state-toricblm-structure-priority-balanced-3day-20260709t185034z-epoch-002
ToricBLM dataset state: toricblm-structure-priority-balanced-3day-20260709T185034Z epoch 002
This dataset repo records the exact local training-data state visible to the dynamic epoch launcher.
It intentionally stores manifests and audit records rather than duplicating large Parquet shards.
Special checkpoint: toricblm-structure-priority-balanced-3day-20260709T185034Z_epoch_002_special_structure_delta_step_002750.pt
Checkpoint repo: AmelieSchreiber/ToricGT_160M_FoT
Curriculum… See the full description on the dataset page: https://huggingface.co/datasets/AmelieSchreiber/toricblm-dataset-state-toricblm-structure-priority-balanced-3day-20260709t185034z-epoch-002.or-bench-balancedSourced from: https://huggingface.co/datasets/bench-llm/or-bench
A balanced dataset of toxic and benign prompts from OR-Bench.
dpo-orpo-mix-38k-balanced
DPO-ORPO-mix-38k
This dataset is intended for use with DPO or ORPO training.
It represents a balanced version of the llmat/dpo-orpo-mix-45k dataset, achieved through a clustering-based approach as outlined in this paper.
The dataset integrates high-quality samples from the following DPO datasets:
argilla/Capybara-Preferences: highly scored chosen answers >=5 (2882 samples)
argilla/distilabel-intel-orca-dpo-pairs: highly scored chosen answers >=9, not in GSM8K (3961 samples)… See the full description on the dataset page: https://huggingface.co/datasets/llmat/dpo-orpo-mix-38k-balanced.balanced-emotion-dataset-majestrino-withtemporal-detailed-captions
Balanced Emotion Dataset — Majestrino with Temporal Detailed Captions
An emotion-balanced subset of TTS-AGI/majestrino-unified-detailed-captions-temporal.
Overview
Total samples: 482,594
Samples per emotion category: 12,997
Number of emotion categories: 40
Format: WebDataset (tar files with FLAC audio + JSON metadata)
Number of tar files: 483
Samples per tar: ~1000
Balancing Strategy
Samples were selected from the source dataset using keyword matching on… See the full description on the dataset page: https://huggingface.co/datasets/TTS-AGI/balanced-emotion-dataset-majestrino-withtemporal-detailed-captions.openhermes116k_balanced
OpenHermes 116K Balanced
This dataset is a balanced 116K-example subset of OpenHermes instruction data.
Method
Examples were grouped by their source label and sampled to reduce source
imbalance. The original conversation structure is preserved.
License
Apache License 2.0.
Disclaimer: Source data can contain errors. Review samples before use.
domain-agnostic-reasoning-traces-balanced-top50-v1
BOTCOIN Balanced Top-50 Reasoning Traces
This public dataset contains enriched BOTCOIN reasoning-trace attempts selected
from canonical dataset/v2 research-ready objects.
Selection policy:
Source only attempts/research-ready objects.
Rank each domain by trace_quality.reasoning_trace_quality_score.
Keep each domain's top 50 percent.
Equalize domains to the smallest top-half count.
The rows are self-contained and intentionally rich: prompt/messages… See the full description on the dataset page: https://huggingface.co/datasets/botcoinmoney/domain-agnostic-reasoning-traces-balanced-top50-v1.nb-asr-numerics-balanced
Balanced Synthetic Norwegian Bokmål Numerics Dataset
This dataset provides a class-balanced synthetic corpus of Norwegian Bokmål sentences containing numeric expressions. It draws 10,000 examples for each of the 59 numeric categories (totaling 590,000 rows).
Source & Synthesis Architecture
Templates source: pere/nb-asr-numerics-categorized.
Methodology:
Filtered the original dataset for kept rows containing annotated entities.
For each target category, sampled 10… See the full description on the dataset page: https://huggingface.co/datasets/pere/nb-asr-numerics-balanced.Qwen3-4B-MATH-traces-balanced
Qwen3-4B MATH Reasoning Traces (Balanced)
Reasoning traces from Qwen/Qwen3-4B on MATH problems, balanced for correct/incorrect.
Model: Qwen/Qwen3-4B (served via vLLM)
Source problems: xDAN2099/lighteval-MATH (train split)
Sampling: Subsampled from the full 10k trace set — 2,500 correct + up to 2,500 incorrect
Generation params: temperature=0.6, top_p=0.95, max_tokens=15000
Problem types: Algebra, Counting & Probability, Geometry, Intermediate Algebra, Number Theory, Prealgebra… See the full description on the dataset page: https://huggingface.co/datasets/jrosseruk/Qwen3-4B-MATH-traces-balanced.cot-statement-qa-broad-v2-balanced
CoT Statement QA (Deterministic)
Conversational supervision dataset for CoT oracles, built from deterministic labels in corpus metadata.
The objective is broad prompt phrasing with high-precision answers.
Data Sources
corpus: data/cot_corpus_v5/corpus_medium.jsonl
importance labels: data/importance_resampled_v2.jsonl
Size
Total rows: 176154
Train: 159320
Validation: 8185
Test: 8649
Task Families
correctness_label: 10000
direct_correctness_label:… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/cot-statement-qa-broad-v2-balanced.open-cai-balanced-partial
Open CAI Balanced Partial
This is a partial generated dataset from the
Open CAI Constitutional AI playground.
It uses prompts and source responses from the harmless-base train split of
Anthropic/hh-rlhf, then
pairs:
a target model's initial response as rejected
a guide-following teacher response as chosen
This snapshot contains 33,711 generated rows. It is not the final full dataset.
Intended Use
This dataset is intended for research on preference modeling… See the full description on the dataset page: https://huggingface.co/datasets/nchapman/open-cai-balanced-partial.balanced-sft-15k
Balanced SFT 15k
Overview
15,000件のSFT用学習データセット。構造化データ生成能力の向上を目的としています。
Component
Samples
Description
Synthetic (template-based)
6,000
Format conversion tasks
daichira/structured-5k-mix-sft
5,000
Extraction/transform tasks
daichira/structured-hard-sft-4k
4,000
Hard tasks
Total
15,000
System Prompt
全サンプルに以下のsystemプロンプトを設定:
You are a precise data transformation assistant.
Output ONLY the requested format.
Do NOT add explanations, comments… See the full description on the dataset page: https://huggingface.co/datasets/Chattso-GPT/balanced-sft-15k.nemo-aot-o3-balanced-expansion-v2
Nemotron AoT-O3 Balanced Expansion V2
Private archive for the unique public-train expansion run. This dataset adds
official train rows that were not present in the previous Tong/DGXChen source,
with two-stage AoT -> AoT-O3 traces for long/search types and deterministic
concise traces for short numeric types.
Expansion rows: 3284
Duplicate rows: {'row_key': 0, 'type_id': 0, 'type_prompt_answer': 0}
{
"bit_manipulation": 210,
"cryptarithm_deduce": 605,
"cryptarithm_guess":… See the full description on the dataset page: https://huggingface.co/datasets/dvyomkesh/nemo-aot-o3-balanced-expansion-v2.turkish-sft-clean-v2-balanced
⚠️ Superseded / Eski Sürüm
Bu repo çalışır durumda olabilir ancak artık önerilen sürüm değildir.
Sebep
Superseded by v4-final. v2 used per-category top-up sources that were found to contain category contamination.
Güncel önerilen dataset
Final Türkçe SFT: https://huggingface.co/datasets/kilicai/turkish-sft-clean-v4-final
Doğrulanmış Türkçe matematik/CoT: https://huggingface.co/datasets/kilicai/turkish-gsm8k-perfect-tr-15k
Yeni eğitimlerde lütfen final temiz… See the full description on the dataset page: https://huggingface.co/datasets/kilicai/turkish-sft-clean-v2-balanced.DeepSeek-R1-Distill-Llama-8B-MATH-traces-balanced
DeepSeek-R1-Distill-Llama-8B MATH Reasoning Traces (Balanced)
4,492 reasoning traces from DeepSeek-R1-Distill-Llama-8B on MATH problems, balanced for correct/incorrect.
Model: deepseek-ai/DeepSeek-R1-Distill-Llama-8B (served via vLLM)
Source problems: xDAN2099/lighteval-MATH (train split)
Sampling: Subsampled from the full 10k trace set — 2,500 correct + 1,992 incorrect (all available incorrect traces)
Generation params: temperature=0.6, top_p=0.95, max_tokens=15000
Accuracy: 55.7%… See the full description on the dataset page: https://huggingface.co/datasets/jrosseruk/DeepSeek-R1-Distill-Llama-8B-MATH-traces-balanced.balanced_smoltalk_basegemma-health-telugu-sft-balanced
Gemma Health Telugu SFT
Splits:
train: 175870 rows
test: 38010 rows
Each row contains:
messages: TRL/Unsloth conversational SFT format.
text: plain serialized chat text fallback.
source, variant, prompt, response: traceability fields.
from datasets import load_dataset
dataset = load_dataset("RohithMidigudla/gemma-health-telugu-sft-balanced", split="train", streaming=True)
test_dataset = load_dataset("RohithMidigudla/gemma-health-telugu-sft-balanced", split="test"… See the full description on the dataset page: https://huggingface.co/datasets/RohithMidigudla/gemma-health-telugu-sft-balanced.Medical-Balanced
Dataset Card for Medical-V2-Balanced
Dataset Summary
Medical-Balanced is a curated dataset focusing on medical text simplification and clinical risk communication. It is "balanced" across different types of medical information—ranging from general health facts to critical surgical warnings.
The goal of this dataset is to train models to communicate medical jargon in a patient-centric way that prioritizes clarity and safety.
Dataset Structure
The dataset is… See the full description on the dataset page: https://huggingface.co/datasets/Akhil-reddy/Medical-Balanced.