datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
sbd-qr-subset
SBD QR Subset — low resolution
A mirror of the low-resolution ROI split of the Synthetic Barcode Dataset
(Quenum, Wang, Zakhor), repackaged from 749,682 loose files into parquet.
split
ROIs
instances
train
80,000
439,731
validation
10,000
55,072
test
10,000
54,876
total
100,000
549,679
Why this repackaging exists
Upstream, this split is three-quarters of a million individual PNG and JPEG
files. That is unpleasant to move, impossible to browse… See the full description on the dataset page: https://huggingface.co/datasets/devmandan/sbd-qr-subset.UPRPRC_SBD_KVchagatai-sbd
Chagatai Sentence Boundary Detection
Canonical word-level Sentence Boundary Detection data for Chagatai. South
Uzbek (uzs) and Uyghur (uig) are optional train-only auxiliary languages.
Every configuration uses the same Chagatai train source split. Validation and
test are physically shared files referenced by all five configurations.
Load with datasets
from datasets import load_dataset
dataset = load_dataset("chagatai-project/chagatai-sbd", "chagatai_only")… See the full description on the dataset page: https://huggingface.co/datasets/chagatai-project/chagatai-sbd.pbe-world
PBE World
A synthetic dataset generator for Programming-by-Examples (PBE) tasks focused on string manipulation pipelines. This is a lightly extended version of PBEBench by Naik et al. 2025.
Overview
PBE World generates tasks where models must reverse-engineer a sequence of string method calls from input/output examples. Each task consists of:
A fixed pipeline of string operations (e.g., s.replace(old, new), s.upper(), s.strip(chars))
Example input/output pairs… See the full description on the dataset page: https://huggingface.co/datasets/sbdzdz/pbe-world.SB_DPO
This dataset is generated by deepseek-ai/DeepSeek-R1-Distill-Qwen-7B with problem set agentica-org/DeepScaleR-Preview-Dataset
For each problem, the model generates 10 completions with temp=0.9 and max_gen_length=8000
Amoung the correct completions, chose the shortest answer as the chosen; Among all completions, choose the longest one as the rejected.
If none of the completions were correct, skip this problem. This means some of the very hard problems are not inluded in this dataset
sbdavijit_sbdh_no_reasoningavijit_sbdhmy-ai-blindspots-assignment
Technical Challenge: Blind Spots of Frontier Models
1. Tested Model
Model Name: SmolLM2-135M
Model Link: HuggingFaceTB/SmolLM2-135M
Parameters: 135 Million (Base Model)
2. Loading Methodology
The model was loaded using the transformers library in a local VS Code environment with the following configuration:
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_id = "HuggingFaceTB/SmolLM2-135M"
tokenizer =… See the full description on the dataset page: https://huggingface.co/datasets/SBD99/my-ai-blindspots-assignment.