datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Puro-2B
Puro-2B Pretraining Data: The Recipe Behind a 2B Model
This is the materialized pretraining data release for
Puro-2B-Base, a 2B base model
trained from scratch on consumer-grade RTX 5090 GPUs.
The repository contains the component-level data pools used to construct the
two Puro-2B pretraining phases, together with the tokenizer used for token
accounting. It is organized for inspection, selective streaming, and recipe
reconstruction rather than as a small train/test… See the full description on the dataset page: https://huggingface.co/datasets/thu-pacman/Puro-2B.PCMind-2.1-Kaiyuan-2B
This repository contains the complete pretraining dataset for
PCMind-v2.1-Kaiyuan-2B, a leading fully open-source language model.
Overview
The dataset is organized into 5 training phases, with all phase datasets open-sourced in this repository. Our training methodology employs domain-specific mixing strategies across five primary domains:
English: General English text
Chinese: General Chinese text
Code: Programming and code-related content
Math: Mathematical reasoning and… See the full description on the dataset page: https://huggingface.co/datasets/thu-pacman/PCMind-2.1-Kaiyuan-2B.PCMind-2.1-Kaiyuan-2B-phase1-part1-1-0323
This repository contains the complete pretraining dataset for
PCMind-v2.1-Kaiyuan-2B, a leading fully open-source language model.
Overview
The dataset is organized into 5 training phases, with all phase datasets open-sourced in this repository. Our training methodology employs domain-specific mixing strategies across five primary domains:
English: General English text
Chinese: General Chinese text
Code: Programming and code-related content
Math: Mathematical reasoning and… See the full description on the dataset page: https://huggingface.co/datasets/Lxd99/PCMind-2.1-Kaiyuan-2B-phase1-part1-1-0323.PCMind-2.1-Kaiyuan-2B-phase1-part1-2
This repository contains the complete pretraining dataset for
PCMind-v2.1-Kaiyuan-2B, a leading fully open-source language model.
Overview
The dataset is organized into 5 training phases, with all phase datasets open-sourced in this repository. Our training methodology employs domain-specific mixing strategies across five primary domains:
English: General English text
Chinese: General Chinese text
Code: Programming and code-related content
Math: Mathematical reasoning and… See the full description on the dataset page: https://huggingface.co/datasets/Lxd99/PCMind-2.1-Kaiyuan-2B-phase1-part1-2.gemma4-e2b-base-topk128-hf-overlay-v128-seed42
Gemma 4 E2B base top-k-128 HF training overlay
This is the immutable training-engine overlay used to distill traces from Gemma 4 E2B base into
Gemma 4 E4B. It preserves the prompts, responses, and exact response token IDs from
JWei05/gemma4-e2b-base-topk128-traces,
but replaces the source vLLM top-k targets with targets recomputed by the Hugging Face training
engine.
This repository is a reproducibility artifact for the corresponding distillation run. It is not a
new… See the full description on the dataset page: https://huggingface.co/datasets/JWei05/gemma4-e2b-base-topk128-hf-overlay-v128-seed42.gollem-corpus-2b-pl
GoLLeM Corpus 2B PL
Dokładny korpus treningowy polskiego modelu bazowego
SlayerLab/GoLLeM-110M-PL-v3
(oraz v2) — ten sam zbiór, po którym model przeszedł dwie epoki. Publikujemy go,
aby każdy mógł odtworzyć trening od zera na własnym tokenizerze.
Jak powstał ten plik. Korpus odzyskano przez zdekodowanie stokenizowanego
checkpointu treningowego (byte-level BPE dynaword-32k, round-trip bezstratny; granice
dokumentów = token <|endoftext|>). To jest dokładnie tekst, który model… See the full description on the dataset page: https://huggingface.co/datasets/SlayerLab/gollem-corpus-2b-pl.qwen3.5-2b-base-blind-spots
Qwen3.5-2B-Base — Blind Spot Analysis (Text + Vision)
Model Tested
Field
Value
Model
Qwen/Qwen3.5-2B-Base
Parameters
2.27 B (2,274 M per HF metadata)
Architecture
Hybrid Gated-DeltaNet (dense FFN) — 24 LM layers (18 DeltaNet + 6 full-attention), ViT vision encoder
Type
Pre-trained base model (not instruction-tuned)
Context
262 144 tokens
Modalities
Text + Vision (early-fusion multimodal)
Key Contributions
Only multimodal… See the full description on the dataset page: https://huggingface.co/datasets/F555/qwen3.5-2b-base-blind-spots.tinybrain-pretrain-corpus-2b
TinyBrain Pretrain Corpus 2B
A mixed-source English pretraining corpus for training small language models.
TinyBrain Pretrain Corpus 2B is a mixed-source dataset built for pretraining small causal language models, especially the TinyBrain-100M Base model.
The dataset combines educational text, factual/wiki-style text, math reasoning data, Python code-summary data, clean web text, and conversation-style data. It is designed to give small models a useful general foundation… See the full description on the dataset page: https://huggingface.co/datasets/exnivo/tinybrain-pretrain-corpus-2b.ml2b
ML2B: Multi-Lingual ML Benchmark For AutoML
This repository provides the dataset for ML2B (Multi-Lingual ML Benchmark for AutoML), the first benchmark for evaluating multilingual machine learning (ML) code generation. Presented in the paper ML2B: Multi-Lingual ML Benchmark For AutoML, ML2B consists of 30 Kaggle competitions translated into 13 natural languages. It covers tabular, text, and image data types, and includes structured metadata and validated human-reviewed… See the full description on the dataset page: https://huggingface.co/datasets/enaix/ml2b.qwen35-2b-tool-use-qwen36-27b-curation-candidates
Full candidate collections: 2B tool use + 27B data curation
This public Dataset contains two complete, unredacted, exact-40 candidate collections:
Tool use: Qwen/Qwen3.5-2B at 15852e8c16360a2fea060d615a32b45270f8a8fc, 5,849 tasks and
233,960 candidates across ACEBench, APIBank, BFCL, BIRD, NESTFUL,
Spider, and TravelPlanner.
Data curation: Qwen/Qwen3.6-27B at 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9, 5,021
targets and 200,840 candidates, plus the source target rows and the… See the full description on the dataset page: https://huggingface.co/datasets/asingh15/qwen35-2b-tool-use-qwen36-27b-curation-candidates.qwen35-2b-personal-training-data
Qwen3.5-2B Three-Domain Training Data
A reproducible training-data release assembled and processed by wisdompan
for Qwen3.5-2B experiments across mathematics, code, and instruction following.
Dataset configurations
Configuration
Purpose
Train rows
Validation rows
full_mix
Unified three-domain student training
86,931
3
teacher_math
Mathematics teacher training
17,917
1
teacher_code
Code teacher training
23,667
1
teacher_if
Instruction-following… See the full description on the dataset page: https://huggingface.co/datasets/wisdompan/qwen35-2b-personal-training-data.Snowball-67B-A2B-RLVR1-Repro-Data
Snowball 67B-A2B RLVR1 data
These are the exact Parquet inputs retained for the Snowball 67B-A2B sync and
async RLVR1 experiments on Iris cw-rno2a in September 2026. The data was
selected from the skyrl_gym route of a TaskTrove conversion of the public
NVIDIA Nemotron RL Ultra training blend,
preserving source order and holding out the last 100 selected rows. See
provenance.json for the local conversion and filtering record. The original
TaskTrove release
is also public.… See the full description on the dataset page: https://huggingface.co/datasets/open-athena/Snowball-67B-A2B-RLVR1-Repro-Data.synthetic-b2b-saas-support-dialogues-sample
Synthetic B2B SaaS Support Dialogues (Sample)
Free sample: 100 dialogues from a larger dataset of 484 synthetic customer support conversations for B2B SaaS products.
What's inside
100 complete dialogues (6–8 messages each)
7 issue categories: auth, billing, integration, data, account, technical, onboarding
Rich metadata: resolution_status, customer_sentiment, agent_actions, escalation_needed
Realistic technical details: error codes, URLs, button names, account… See the full description on the dataset page: https://huggingface.co/datasets/Jurgen1161/synthetic-b2b-saas-support-dialogues-sample.Taiwan-Text-Excellence-2B
High quality corpus for Taiwanese culture and Traditional Chinese
Taiwan Text Excellence (TTE)
Contains high quality news and articles in Traditional Chinese.
The data processing pipeline is optimized for LLM performance.
Is de-duplicated and cleaned using both rule-based and learning-based filters.
E.g., urls/emails/html tags/abnormal characters are cleaned, and numbers (full-width or half-width) are normalized.
Contains ~2 billion tokens, measured using BPE tokenizer… See the full description on the dataset page: https://huggingface.co/datasets/liswei/Taiwan-Text-Excellence-2B.caliber-extension-gemma4-e2b-grpo-rollouts
CALIBER Extension — Gemma4-E2B GRPO Rollouts
Training rollouts from matched GRPO arms on google/gemma-4-E2B-it
(new-prompt template, non-thinking, full bf16, max completion 1500, 150 steps).
Subsets
subset
arm
τ
prior
rows
mean reward_total
accuracy
full schema
caliber
vanilla CALIBER
0.0
—
1600
2.298
0.514
0.664
mink
Min-K% prior
1.0
mink_0.2
4800
2.506
0.520
0.680
minkpp
Min-K++% prior
1.0
minkpp_0.2
4800
2.637
0.541
0.726
Load:
from datasets… See the full description on the dataset page: https://huggingface.co/datasets/dmnsh/caliber-extension-gemma4-e2b-grpo-rollouts.synthetic_polistance
Fully Synthetic Prompts for LLM Political Stance Detection
All resources developed in the article "Templated or fully Synthetic? Prompt construction as a confound in measuring LLM political stance beyond writing assistance" (Chalkidis, 2026).
Paper Abstract
Political stance detection in LLMs has long been dominated by closed-ended, multiple-choice political survey questions—originally designed for humans, and thus lacks the realism and nuance of human-AI… See the full description on the dataset page: https://huggingface.co/datasets/kiddothe2b/synthetic_polistance.nla-gemma4e2b-relabel-v1-eval
Gemma-4-E2B layer-23 evaluation set, relabeled (v1), with contamination flags
The 580-document evaluation pool on which every activation-verbalizer result in this
project is scored, with each row's evaluation text rewritten from a topic summary to
a feature-attribution label. The activations are byte-identical to the original
evaluation set; only the text column changed, and the original text is preserved.
This pool is not disjoint from the training corpus. Read this… See the full description on the dataset page: https://huggingface.co/datasets/Solshine/nla-gemma4e2b-relabel-v1-eval.saferide-gemma-4-e2b-v058-original-419806-training-data
SafeRide Synthetic Bilingual Safety Guidance Dataset v0.5.8
This research and development dataset contains synthetic English and Kiswahili
chat conversations. It was designed to help a language model practice cautious,
agency-preserving safety guidance, useful refusal behavior, and responses that
avoid inventing facts. It contains no real survivor reports or production
records. The frozen dataset is publicly available under Creative Commons
Attribution 4.0 International (CC BY… See the full description on the dataset page: https://huggingface.co/datasets/esherialabs/saferide-gemma-4-e2b-v058-original-419806-training-data.qwen35-2b-tool-use-candidates
Qwen3.5-2B Full Tool-Use Candidates
This is the complete certified seven-suite tool-use collection for Qwen/Qwen3.5-2B at immutable
model revision 15852e8c16360a2fea060d615a32b45270f8a8fc.
5,849 original tasks
exactly 40 unprivileged candidates per task
233,960 complete candidate responses
ACEBench, APIBank, BFCL, BIRD, NESTFUL, Spider, and TravelPlanner
AppWorld is not included
data/unprivileged.jsonl is a byte-for-byte copy of the certified collection. Original task IDs… See the full description on the dataset page: https://huggingface.co/datasets/asingh15/qwen35-2b-tool-use-candidates.nla-gemma4e2b-relabel-v1-corpus
Gemma-4-E2B layer-23 activation corpus, relabeled (v1)
1356 training rows for an activation verbalizer. Each row pairs a residual-stream
activation captured at layer 23 of google/gemma-4-E2B with a natural-language label
describing what the model must have integrated at that position to predict its next
token. This is the training set behind
Solshine/gemma-4-e2b-nla-L23-av-priordev-relabel-v1-wd3.
Why it exists
An audit of the previous version of this corpus found… See the full description on the dataset page: https://huggingface.co/datasets/Solshine/nla-gemma4e2b-relabel-v1-corpus.Gemma-4-E2B-SSFT
Gemma-4-E2B-SSFT
This is a dataset that has been created using SSFT (simple-sft), a synthetic data generation tool written by me.
It contains a few hundred samples for testing.
Try it out yourself!
qwen3.5-2B-vi-query
Vietnamese Medical Query Normalization / Expansion / Routing Pack (v3)
1162 synthetic ChatML examples for fine-tuning a small Vietnamese model (target: Qwen/Qwen3.5-2B,
trained with Unsloth) to turn a raw, everyday Vietnamese medical query into structured JSON:
normalized query, intent, entities, must-preserve tokens, lexical/semantic query variants, and a
retrieval-routing hint, for a downstream medical RAG system.
The model does not answer medical questions. It only normalizes… See the full description on the dataset page: https://huggingface.co/datasets/daipham31/qwen3.5-2B-vi-query.Pwen3.5_2B_Python_Finetune
Pwen3.5-2B-Coding-Finetune
Pwen 3.5 2B Coding Dataset
A high-quality instruction dataset for fine-tuningQwen3.5-2B into a concise coding assistant
Created by Pavel Hanzel
Overview
Pwen3.5-2B-Coding-Finetune is an instruction tuning dataset designed to transform Qwen3.5-2B into a practical programming assistant.
The dataset focuses on:
Python programming
Debugging
Code explanations
Development workflows
AI/LLM usage
Direct technical… See the full description on the dataset page: https://huggingface.co/datasets/FreeAIn/Pwen3.5_2B_Python_Finetune.ai-vs-human-google-gemma-2-2b-it
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-google-gemma-2-2b-it.ultrachat_200k_generated_gemma-2-2b-itThis dataset contains 512 answers generated by the gemma-2-2b-it model on a subset of the ultrachat 200k test_sft dataset using greedy decoding.
The subset was generated by filtering out conversations that were >= 1024 - 128 tokens long, and answers were cut off at each batch after 1024 - min(batch_prompt_lengths) generated tokens, such that each answer is at most 128 tokens long. The generated answers are 200k tokens so 390 tokens (~300 words or 2/3 pages) on average.
youtu-llm-2b-base-blind-spots
Youtu-LLM-2B-Base Blind Spots Evaluation Dataset
This dataset contains 75 evaluation prompts used to analyze the failure modes of tencent/Youtu-LLM-2B-Base,
a 1.96B parameter dense base language model released on December 31, 2025. Each row includes the input prompt, the expected answer, and the model’s
generated output obtained during inference on a Google Colab T4 GPU.
The prompts span 13 broad categories including arithmetic, logic, multilingual generation, instruction following… See the full description on the dataset page: https://huggingface.co/datasets/k-imtz/youtu-llm-2b-base-blind-spots.gemma-4-e2b-deception-behavior-completions
Gemma-4-E2B deception & behavior completions
Consolidated 910-row corpus of (scenario prompt + Gemma-4-E2B-generated completion) pairs from earlier mechanistic-interpretability experiments. Each row captures the prompt the model saw and the text it actually produced; for a subset, Claude-Haiku-4-5 judge verdicts and SAE-feature labels are included.
The corpus is meant to be used as activation-extraction input for downstream interpretability work — Natural Language Autoencoder (NLA)… See the full description on the dataset page: https://huggingface.co/datasets/Solshine/gemma-4-e2b-deception-behavior-completions.french-philosophy-10K
Philosophy
Langue Française
Dataset de Pre-Training
Ce jeu de données propose 10 000 exemples soigneusement rédigés en français, représentant environ 1,2 million de jetons. Il est destiné spécifiquement au pré-entraînement ou au fine-tuning de… See the full description on the dataset page: https://huggingface.co/datasets/Dorian2B/french-philosophy-10K.gemma-4-e2b-nla-av_sft-v0_1_x-gemini-persona-audit
Gemma-4-E2B NLA AV-SFT Training Corpus (v0.1.x, Gemini persona+audit)
The 4,734-row AV-SFT training corpus for the v0.1.x Gemma-4-E2B NLA — a 9-source-family diversified expansion over the v0.0.x OpenWebText-only corpus. Labels generated by Gemini CLI following the persona+audit pipeline (Dr. Marisol Chen labels, Dr. Riley Otsuka audits).
This is the in-progress v0.1.x labeled training set. AR-SFT companion is still being labeled (~16% complete as of this dataset publish). When the… See the full description on the dataset page: https://huggingface.co/datasets/Solshine/gemma-4-e2b-nla-av_sft-v0_1_x-gemini-persona-audit.gemma-2-2b-blind-spots
Gemma-2-2B Base Model Blind Spots Dataset
Dataset Description
This dataset contains 10 carefully curated examples that highlight specific blind spots and failure modes of the google/gemma-2-2b base model. The examples span diverse categories of reasoning and computation where the base model demonstrates systematic weaknesses.
Model Tested: google/gemma-2-2b
Type: Base model (pre-trained, not instruction-tuned)
Parameters: 2.6B
Release Date: 2024
Methodology… See the full description on the dataset page: https://huggingface.co/datasets/SumaiyaMifra/gemma-2-2b-blind-spots.