hrm
Datasets
All datasets matching “hrm”HRM-He-corpus-objective
Hebrew reasoning traces
Generated Hebrew chain-of-thought over code, cybersecurity, agentic, math and
general-reasoning seeds. Built for a Hebrew/English code-specialised LM, where
off-the-shelf Hebrew reasoning data is effectively nonexistent.
What the default config contains
Every row the training corpus keeps -- not a filtered highlight reel. Two things
are disqualifying and are absent: a wrong final answer (answer_ok is False), and
Arabic drift. Everything… See the full description on the dataset page: https://huggingface.co/datasets/guychuk/HRM-He-corpus-objective.HRM-Text-data-io-cleaned-20260515Pre-built HRM-Text pretraining dataset from raw data using the data_io cleaning scripts.
Citation
If you find this project or our paper useful, please consider citing our paper:
@misc{wang2026hrmtextefficientpretrainingscaling,
title={HRM-Text: Efficient Pretraining Beyond Scaling},
author={Guan Wang and Changling Liu and Chenyu Wang and Cai Zhou and Yuhao Sun and Yifei Wu and Shuai Zhen and Luca Scimeca and Yasin Abbasi Yadkori},
year={2026}… See the full description on the dataset page: https://huggingface.co/datasets/sapientinc/HRM-Text-data-io-cleaned-20260515.hrm-tokenized-bpe65khebrew-hrm-corpus
Hebrew HRM-Text Corpus
Training corpus for a Hebrew Hierarchical Reasoning Model, replicating the
sapientinc/HRM-Text-1B recipe
(train-from-scratch, PrefixLM over {condition, instruction, response}, loss on response only).
Schema
Each line: {"condition": "<tags>", "instruction": "...", "response": "..."}.
Condition tags map to special tokens: direct→<|object_ref_start|>, cot→<|object_ref_end|>,
noisy→<|quad_start|>, synth→<|quad_end|> (composite tags… See the full description on the dataset page: https://huggingface.co/datasets/guychuk/hebrew-hrm-corpus.HRM8K
| 📖 Paper | 📝 Blog | 🖥️ Code(Coming soon!) |
HRM8K
We introduce HAE-RAE Math 8K (HRM8K), a bilingual math reasoning benchmark for Korean and English.
HRM8K comprises 8,011 instances for evaluation, sourced through a combination of translations from established English benchmarks (e.g., GSM8K, MATH, OmniMath, MMMLU) and original problems curated from existing Korean math exams.
Benchmark Overview
The HRM8K benchmark consists of two subsets:
Korean School Math (KSM):… See the full description on the dataset page: https://huggingface.co/datasets/HAERAE-HUB/HRM8K.HR-MMSearch
Dataset Description
HR-MMSearch is a benchmark designed to evaluate the Agentic Reasoning and Search capabilities of Multimodal Large Language Models in complex visual tasks.
This dataset was introduced by SenseTime Research in the paper SenseNova-MARS: Empowering Multimodal Agentic Reasoning and Search via Reinforcement Learning.
Key Features:
High-Resolution Images: Contains high-resolution image inputs, requiring the model to possess fine-grained visual perception… See the full description on the dataset page: https://huggingface.co/datasets/sensenova/HR-MMSearch.