datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
agentic-llm-pretraining-1.7b
Agentic LLM Pretraining Dataset
A pretraining corpus for small language models (1-3B parameters) optimized for agentic tasks. The corpus emphasizes learning to comprehend language, reason, follow instructions, and use tools over memorizing factual knowledge — the assumption is that domain knowledge will be provided at runtime via RAG. The idea is that this could enable much smaller pretraining corpora by omitting the large volumes of text typically needed to memorize facts.… See the full description on the dataset page: https://huggingface.co/datasets/visionscaper/agentic-llm-pretraining-1.7b.ai-vs-human-HuggingFaceTB-SmolLM2-1.7B-Instruct
AI vs Human dataset on the CNN Daily mails
Dataset Description
This dataset showcases pairs of truncated articles and their respective completions, crafted either by humans or an AI language model.
Each article was randomly truncated between 25% and 50% of its length.
The language model was then tasked with generating a completion that mirrored the characters count of the original human-written continuation.
Data Fields
'human': The original human-authored… See the full description on the dataset page: https://huggingface.co/datasets/zcamz/ai-vs-human-HuggingFaceTB-SmolLM2-1.7B-Instruct.Qwen3-1.7B-UHead-Rollouts
Qwen3-1.7B UHead Rollouts
Reusable rollouts and UHead-derived supervision produced by the Qwen3-1.7B
native-thinking experiments on CSCS Clariden in August 2026. The release keeps
token-level provenance and decoded text while excluding model checkpoints,
W&B files, logs, repeated guard inputs, failed smoke runs, and duplicated
generation artifacts.
Configurations
natural_chains (9,376 rows)
Natural Qwen3-1.7B reasoning chains over 600 math… See the full description on the dataset page: https://huggingface.co/datasets/JingweiNi/Qwen3-1.7B-UHead-Rollouts.dapo-math-17k-qwen3-1.7b-base-n8
DAPO-Math-17k sampled with Qwen3-1.7B-Base, n=8
17398 problems from the RL training set, each sampled 8 times and scored with
the reward function the RL runs themselves used.
The point of this dataset is to be comparable with what the RL runs actually saw, so
every sampling knob is taken from the live training config or from the default that
config falls through to. Two of them are not in the config file at all and would be
wrong if guessed: top_k = -1 and min_tokens = 1.… See the full description on the dataset page: https://huggingface.co/datasets/RyanYr/dapo-math-17k-qwen3-1.7b-base-n8.RLVE-Qwen3-1.7B-Pass1-Rollouts
RLVE teacher rollouts — Qwen3-1.7B (pass@1)
Teacher rollouts for on-policy distillation on the RLVE environment suite.
Teacher / sampler: Qwen3-1.7B
Source prompts: RLVE train split — 9000 questions across RLVE-Eval Gym
environments (counting / combinatorics / optimization tasks)
Sampling: 1 sample/question (pass@1) = 9000 records,
temperature 0.7, max 4096 new tokens
Rewards: inline RLVE-Eval Gym verifier score (continuous, in [-1, 1]).
Teacher accuracy (reward>0): 20 / 9000 =… See the full description on the dataset page: https://huggingface.co/datasets/CL-From-Nothing/RLVE-Qwen3-1.7B-Pass1-Rollouts.qwen3-1.7b-traces
Qwen3-1.7B reasoning traces
On-policy chain-of-thought rollouts from Qwen/Qwen3-1.7B
in thinking mode, collected to study long-context KV-cache residency constraints (trainable
sparse attention, in the InfLLM-V2 /
NOSA line).
Used as adaptation data: training a sparse+local attention variant on the model's own output
distribution avoids the alignment tax that continued pretraining on external corpora imposes on a
post-trained model.
Caveats
Not filtered for… See the full description on the dataset page: https://huggingface.co/datasets/kmoss/qwen3-1.7b-traces.Qwen-3-1.7B-with-Reasoning-x500
Qwen-3-1.7B-with-Reasoning-x500
This is version v1 - we continue updating and upscaling this dataset!
Overview
This is a high-quality synthetic dataset consisting of 500 diverse samples generated by Qwen 3 1.7B.
The goal of this dataset is to provide clean, direct, and logical reasoning traces for distilling larger model capabilities into Small Language Models (SLMs) like my Apex models or those of CompactAI.
Dataset Structure
The data is provided… See the full description on the dataset page: https://huggingface.co/datasets/LH-Tech-AI/Qwen-3-1.7B-with-Reasoning-x500.DeepScaleR-Qwen3-1.7B-2k-diverse-agreed-coded
DeepScaleR-Qwen3-1.7B-2k diverse-agreed, strategy-coded
1635 competition-math problems (the claude_agrees_gold == True subset of a
2k diverse-classified DeepScaleR pool). Each row carries Claude's worked
claude_solution plus three leak-free re-expressions of the strategy it
deploys, drawn from a shared 116-code strategy codebook.
Columns
idx — row index into agentica-org/DeepScaleR-Preview-Dataset (resume/join key).
problem, answer — the problem and gold answer.… See the full description on the dataset page: https://huggingface.co/datasets/zjhhhh/DeepScaleR-Qwen3-1.7B-2k-diverse-agreed-coded.agentic-llm-pretraining-1.7b-tokenized-qwen3-4k
Agentic LLM Pretraining Dataset - Tokenized (Qwen3, 4K context)
Pre-tokenized version of visionscaper/agentic-llm-pretraining-1.7b for pre-training small language models for agentic AI use cases.
Overview
Property
Value
Source dataset
visionscaper/agentic-llm-pretraining-1.7b
Tokenizer
Qwen/Qwen3-1.7B
Context length
4,096 tokens
EOD token
<|endoftext|> (ID 151643)
Token dtype
uint32
Total samples
375,384
Total tokens
~1.54 billion
Storage
~5.8 GB… See the full description on the dataset page: https://huggingface.co/datasets/visionscaper/agentic-llm-pretraining-1.7b-tokenized-qwen3-4k.Qwen3-1.7B_Legal_Distill
⚖️ Qwen3-1.7B Legal Distill Dataset
This repository contains a specialized dataset designed for Knowledge Distillation and Policy Refinement, generated during a Group Relative Policy Optimization (GRPO) training run. 🚀
The core philosophy of this dataset is simultaneous logging and distillation: rather than discarding the multiple generation outputs produced during the RLHF/GRPO process, we capture and log them alongside their respective reward scores. This creates a rich… See the full description on the dataset page: https://huggingface.co/datasets/Azzindani/Qwen3-1.7B_Legal_Distill.Qwen-3-1.7B-with-Reasoning-x100
Qwen-3-1.7B-with-Reasoning-x100
This is version v1 - we continue updating and upscaling this dataset!
Overview
This is a high-quality synthetic dataset consisting of 100 diverse samples generated by Qwen 3 1.7B.
The goal of this dataset is to provide clean, direct, and logical reasoning traces for distilling larger model capabilities into Small Language Models (SLMs) like my Apex models or those of CompactAI.
Dataset Structure
The data is provided… See the full description on the dataset page: https://huggingface.co/datasets/LH-Tech-AI/Qwen-3-1.7B-with-Reasoning-x100.rose_code-Qwen3-1.7B-Pass8-Rollouts
rose_code rollouts — Qwen3-1.7B (pass@8)
Model rollouts on the rose_code test split, for the OPD coding pipeline.
Model / sampler: Qwen3-1.7B
Source prompts: CL-From-Nothing/rose_code test split — 408 competitive-programming questions (codeforces-style)
Sampling: 8 samples/question (pass@8) = 3264 records, temperature 0.7, max 16384 new tokens, max_model_len 32000
Rewards: DeepCoder code verifier (deepcoder_reward_fn.py) — 1.0 if the generated program passes all unit tests… See the full description on the dataset page: https://huggingface.co/datasets/CL-From-Nothing/rose_code-Qwen3-1.7B-Pass8-Rollouts.qwen3-1.7b-blind-spots
Blind Spots of a Frontier Base Model: Failure Analysis of Qwen3-1.7B-Base
Dataset Link
Public Hugging Face Dataset
1. Model Selection
To complete this challenge, I browsed recently released open models on
Hugging Face within the 0.6B--6B parameter range.
The model selected for analysis:
Model Name: Qwen3-1.7B-Base Parameter Size: 1.7B Modality: Text (Causal Language Model) Type: Base model (not instruction-tuned) Availability: Public on Hugging Face… See the full description on the dataset page: https://huggingface.co/datasets/Habibgm/qwen3-1.7b-blind-spots.whetstone-Qwen-1.7B-generations
whetstone-Qwen-1.7B-generations
Paired verbose and compact-register reasoning traces for 2,414 maths
problems, with per-trace follow-ability scores.
Each row holds a problem, the long chain-of-thought Qwen3-1.7B produced for it,
a compact-notation rewrite of that same reasoning, token counts for both, and
the scores used to measure how followable the compact version is to Qwen3-1.7B.
11,174,460 original think tokens → 750,087 compressed (14.9×).
Selection
Every… See the full description on the dataset page: https://huggingface.co/datasets/thebajajra/whetstone-Qwen-1.7B-generations.DeepScaleR-Qwen3-1.7B-2k-strategy-error-200
DeepScaleR Qwen3 1.7B 2K strategy errors
This dataset contains 200 distinct questions selected from
zjhhhh/DeepScaleR-Qwen3-1.7B-2k-agreed-regraded-le5-coded
at revision 8b6e0f481bced00132c95fb631745d4992fa19fd.
Each row has one manually selected model response whose main failure is a
strategy error relative to the source row's code_hint: the response does not
materially use the hint's core route, substitutes another strategy, or omits a
decisive hinted stage in favor of an… See the full description on the dataset page: https://huggingface.co/datasets/zjhhhh/DeepScaleR-Qwen3-1.7B-2k-strategy-error-200.rlve_offline_20K_POPE_prefix_pass1_qwen3-1.7b
RLVE offline-20K POPE-prefix completions — Qwen3-1.7B (pass1)
Prefix-conditioned completions generated by Qwen3-1.7B over the
rlve_offline_20K_POPE_prefix prompt set (20000 records, 1 sample/prompt).
Produced by SLURM job 6580578 (vLLM, tp=2), 2026-06-15.
Fields
index, sample_id, prompt, prefix, response, answer, rewards
⚠️ Caveat on rewards
The inline rewards field is all 0.0 — this is the known inline-Gym-verifier
artifact (same as the old… See the full description on the dataset page: https://huggingface.co/datasets/CL-From-Nothing/rlve_offline_20K_POPE_prefix_pass1_qwen3-1.7b.smollm2-1.7b-blind-spots
SmolLM2-1.7B Blind Spots Dataset
A curated evaluation dataset documenting specific failure modes of HuggingFaceTB/SmolLM2-1.7B — a 1.7 billion parameter pretrained (base) language model. Each entry contains a completion-style prompt, the verified correct answer, and the model's actual incorrect output produced via deterministic greedy decoding.
This dataset was created as part of the "Blind Spots of Frontier Models" technical challenge to systematically identify where small… See the full description on the dataset page: https://huggingface.co/datasets/Shinzmann/smollm2-1.7b-blind-spots.local-code-arena-mbpp-qwen3_1.7b
Local Code Arena Telemetry: MBPP Benchmark on Qwen3 1.7B
This repository hosts the raw evaluation metrics, execution telemetry logs, and structural syntax outputs captured from running the Mostly Basic Python Problems (MBPP) benchmark against the next-generation Qwen3 1.7B parameter model.
This specific run marks our transition into evaluating generalist next-generation architectures on local consumer hardware, providing a direct point of comparison against older, specialized… See the full description on the dataset page: https://huggingface.co/datasets/ShahzebKhoso/local-code-arena-mbpp-qwen3_1.7b.RLVE-Test-Qwen3-1.7B-GRPO-step70-Pass8
RLVE test eval — GRPO step70 (pass@8)
Evaluation rollouts on the RLVE test split.
Model: grpo_train_Qwen3-1.7B-SFT-rlve-20K-1epoch (GRPO, step 70)
Source prompts: RLVE test split — 180 questions (RLVE-Eval Gym environments)
Sampling: 8 samples/question (pass@8) = 1440 records, temperature 0.7,
max 16384 new tokens
Rewards: inline RLVE-Eval Gym verifier score (continuous, in [-1, 1]).
Record-level accuracy (reward>0): 216 / 1440 = 15.0%, mean reward -0.677
pass@8 (>=1 of 8… See the full description on the dataset page: https://huggingface.co/datasets/CL-From-Nothing/RLVE-Test-Qwen3-1.7B-GRPO-step70-Pass8.qwen3-1.7b-base-blind-spots
Qwen3-1.7B-Base Blind Spots Dataset
Model Tested
Model: surogate/Qwen3-1.7B-Base
Architecture: Qwen3 (decoder-only transformer)
Parameters: 1,720,574,976 (~1.7B)
Type: Base pretrained model (NOT instruction-tuned)
License: Apache 2.0
Release: May 2025
How the Model Was Loaded
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_id = "surogate/Qwen3-1.7B-Base"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model =… See the full description on the dataset page: https://huggingface.co/datasets/RHYTHM1028/qwen3-1.7b-base-blind-spots.smollm2-1.7b-blindspots
SmolLM2-1.7B Blind Spots Dataset
A curated dataset of 12 diverse probe examples where the base language model
HuggingFaceTB/SmolLM2-1.7B
makes incorrect or unreliable predictions. Each row contains the raw prompt, the
expected correct answer, the model's actual output (greedy decoding), the error
category, and an explanation.
Tested Model
HuggingFaceTB/SmolLM2-1.7B
Property
Value
Parameters
1.7 billion
Type
Pure base model (pretrained only — no… See the full description on the dataset page: https://huggingface.co/datasets/Shah-4-8-1-2/smollm2-1.7b-blindspots.ThinkSafe-Qwen3-1.7B-WildGuard
ThinkSafe Dataset
This dataset is associated with the paper THINKSAFE: Self-Generated Safety Alignment for Reasoning Models.
Paper: https://arxiv.org/abs/2601.23143GitHub: https://github.com/seanie12/ThinkSafe.git
Dataset Description
This dataset contains 39,787 training examples with instructions and responses, labeled for safety alignment. Each example includes:
instruction: The input instruction
response: The generated response
prompt_label: Safety label for the… See the full description on the dataset page: https://huggingface.co/datasets/Seanie-lee/ThinkSafe-Qwen3-1.7B-WildGuard.agentic-llm-pretraining-1.7b
Agentic LLM Pretraining Dataset
A pretraining corpus for small language models (1-3B parameters) optimized for agentic tasks. The corpus emphasizes learning to comprehend language, reason, follow instructions, and use tools over memorizing factual knowledge — the assumption is that domain knowledge will be provided at runtime via RAG. The idea is that this could enable much smaller pretraining corpora by omitting the large volumes of text typically needed to memorize facts.… See the full description on the dataset page: https://huggingface.co/datasets/travisp83/agentic-llm-pretraining-1.7b.ultrachat-qwen3-1.7b-think
UltraChat reasoning traces — Qwen3-1.7B
Chain-of-thought traces written by Qwen3-1.7B 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": "73",
"teacher": "Qwen3-1.7B",
"problem": "When it comes to burgers, do you prefer beef, chicken… See the full description on the dataset page: https://huggingface.co/datasets/VertexResearch/ultrachat-qwen3-1.7b-think.RLVE-Eval20-Qwen3-1.7B-SSD-N20-SFT-Train
RLVE-Eval20-Qwen3-1.7B-SSD-N20-SFT-Train
Parquet for supervised fine-tuning (SFT) in a Simple Self-Distillation (SSD) style setup on the RLVE eval20 filtered pool.
What this is
Questions: 800 prompts from rlve_eval20_filtered (20 harder RLVE environments × 40 seeds).
Continuations: For each question, N = 20 completions were sampled from a frozen Qwen3-1.7B student (including the initial greedy / first sample and additional stochastic samples), then packed for SFT.
Rows:… See the full description on the dataset page: https://huggingface.co/datasets/CL-From-Nothing/RLVE-Eval20-Qwen3-1.7B-SSD-N20-SFT-Train.RLVE-Test-Qwen3-1.7B-SFT-warmup-Pass8
RLVE test eval — SFT warmup (pass@8)
Evaluation rollouts on the RLVE test split.
Model: Qwen3-1.7B-SFT-rlve-20K-1epoch (RL-warmup baseline, pre-RL)
Source prompts: RLVE test split — 180 questions (RLVE-Eval Gym environments)
Sampling: 8 samples/question (pass@8) = 1440 records, temperature 0.7,
max 16384 new tokens
Rewards: inline RLVE-Eval Gym verifier score (continuous, in [-1, 1]).
Record-level accuracy (reward>0): 176 / 1440 = 12.2%, mean reward -0.724
pass@8 (>=1 of 8… See the full description on the dataset page: https://huggingface.co/datasets/CL-From-Nothing/RLVE-Test-Qwen3-1.7B-SFT-warmup-Pass8.