datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
gsm8k-qwen3-0.6B-rollouts
GSM8K Qwen3-0.6B Rollouts
Verifier-labeled solutions sampled from Qwen/Qwen3-0.6B with vLLM
0.26.0 and its pytorch top-k/top-p
sampler for every prompt in the train and test splits of
openai/gsm8k.
Dataset size and label balance
train: 7,473 prompt rows; 747,300 rollouts; 602,611 correct (80.64%), 144,689 incorrect (19.36%)
test: 1,319 prompt rows; 131,900 rollouts; 102,245 correct (77.52%), 29,655 incorrect (22.48%)
Each row contains one source prompt and 100… See the full description on the dataset page: https://huggingface.co/datasets/RLHF-Book/gsm8k-qwen3-0.6B-rollouts.Qwen3-0.6B-cybertown-RLVR-data
Qwen3-0.6B-cybertown-RLVR-data
This dataset contains Cybertown RLVR training and validation data used for WindyLab/Qwen3-0.6B-cybertown-RLVR.
Files
train.parquet: RLVR training split.
val.parquet: RLVR validation split.
manifest.json: dataset construction metadata.
states.index.json: index mapping state ids to sharded state records.
states.shards.manifest.json: state shard metadata.
states/: sharded replan state records.
The legacy monolithic states.jsonl is… See the full description on the dataset page: https://huggingface.co/datasets/WindyLab/Qwen3-0.6B-cybertown-RLVR-data.clean-gsm8k-aug-qwen3-0.6b
Clean GSM8K-Aug Qwen3-0.6B
This dataset replaces the reasoning steps and answers in
cs-giung/clean-gsm8k-aug
with responses generated by
Qwen/Qwen3-0.6B.
The questions and split ordering match source revision
60f9c039ae300041b5dca5dc1482c8ff1ef1eb47. The model revision is
c1899de289a04d12100db370d81485cdf75e47ca.
Generation
Qwen3 thinking mode was enabled. Sampling used:
temperature: 0.6
top-p: 0.95
top-k: 20
min-p: 0.0
maximum new tokens: 32768
initial seed: 0… See the full description on the dataset page: https://huggingface.co/datasets/cs-giung/clean-gsm8k-aug-qwen3-0.6b.ThinkSafe-Qwen3-0.6B-WildGuard
THINKSAFE Dataset
This dataset is part of the THINKSAFE: Self-Generated Safety Alignment for Reasoning Models project.
Dataset Description
This dataset contains safety-aligned training data for reasoning models, specifically generated using the Qwen3-0.6B model with WildGuard safety evaluation. It includes instructions, responses, and safety labels for both prompts and responses.
Dataset Structure
The dataset contains 39,949 examples with the following fields:… See the full description on the dataset page: https://huggingface.co/datasets/Seanie-lee/ThinkSafe-Qwen3-0.6B-WildGuard.countdown-qwen3-0.6b
Countdown Qwen3-0.6B Pass@10 Buckets
Countdown arithmetic problems filtered by observed local Qwen/Qwen3-0.6B success rate over 10 rollouts per problem.
Each problem asks for an arithmetic expression that reaches a target using each listed source number at most once. The final answer should be inside \boxed{...}. Canonical solutions are provided, but any verifier-valid expression is accepted.
Subsets
subset
source bucket
count
observed successes out of 10… See the full description on the dataset page: https://huggingface.co/datasets/simpissa/countdown-qwen3-0.6b.qwen3-0.6B-interleaved-thinking-data
Qwen3 0.6B Interleaved Thinking Data
This dataset contains 8,704 pretraining-style text chunks augmented with short interleaved teacher thoughts. It was built for the blog post Self-Improving Pretraining as a Substrate for Agentic Post-Training.
The dataset turns ordinary pretraining text into the supervised stage of a thinking mid-training pipeline. A teacher inserts short local thoughts into raw FineWeb-Edu chunks while preserving the original text. The student then learns the… See the full description on the dataset page: https://huggingface.co/datasets/Jarrodbarnes/qwen3-0.6B-interleaved-thinking-data.qwen3-0.6b-base-blind-spots
Qwen3-0.6B-Base Blind Spots Dataset
Model Tested
Qwen/Qwen3-0.6B-Base — a 0.6B-parameter
causal language model (pretrained only, no instruction tuning), released May 2025
under the Apache 2.0 license.
How the Model Was Loaded
The model was loaded on Google Colab (T4 GPU, free tier) using Hugging Face
transformers >= 4.51.0:
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
MODEL_NAME = "Qwen/Qwen3-0.6B-Base"
tokenizer =… See the full description on the dataset page: https://huggingface.co/datasets/bechir09/qwen3-0.6b-base-blind-spots.ultrachat-qwen3-0.6b-think
UltraChat reasoning traces — Qwen3-0.6B
Chain-of-thought traces written by Qwen3-0.6B with thinking enabled, answering questions from
UltraChat. Each row pairs a question with the
model's full <think>…</think> reasoning and the answer that followed it.
Intended as training data for distilling reasoning behaviour and chat format into small models.
A row
{
"id": "12",
"teacher": "Qwen3-0.6B",
"problem": "How does this textbook compare to other textbooks on the… See the full description on the dataset page: https://huggingface.co/datasets/VertexResearch/ultrachat-qwen3-0.6b-think.Qwen3-0.6B-AIME-2023-2024-2025-sampling64
Model : Qwen3-0.6B
Original Dataset : AIME2023, AIME2024, AIME2025
Prompt:
{"role": "user", "content": "Please reason step by step, and put your final answer within \boxed{}." + '\n\n' + problem}
Sampling Parameters :
num_sampling=64
max_tokens=38912
temperature=0.6
top_p=0.95
top_k=20
min_p=0
‘correct’ : computed by the code in the link (https://github.com/LeapLabTHU/Absolute-Zero-Reasoner/blob/master/absolute_zero_reasoner/rewards/math_utils.py)
Qwen3-0.6B-MATH-sampling64
Model : Qwen3-0.6B
Original Dataset : MATH
train : first 500 queries in training split
test : MATH500
Prompt:
{"role": "user", "content": "Please reason step by step, and put your final answer within \boxed{}." + '\n\n' + problem}
Sampling Parameters :
num_sampling=64
max_tokens=38912
temperature=0.6
top_p=0.95
top_k=20
min_p=0
‘correct’ : computed by the code in the link (https://github.com/LeapLabTHU/Absolute-Zero-Reasoner/blob/master/absolute_zero_reasoner/rewards/math_utils.py)
lora-rules-qwen3-0.6b-r8-n180
LoRA Rules Dataset
Synthetic behavioral rules dataset for training a hypernetwork that generates
LoRA adapters on-the-fly from structured rule strings.
Format
Each record is a JSON line with fields:
rule_id — unique identifier
rule_type — one of: Constraint, Format, Knowledge, Persona, Safety, Tone
weight — float 0.0–1.0, importance of the rule
description — natural language rule description
raw — full rule string [RuleType|Weight] Description
training_examples — list of… See the full description on the dataset page: https://huggingface.co/datasets/broadfield-dev/lora-rules-qwen3-0.6b-r8-n180.