ISR
Datasets
All datasets matching “ISR”ProverbEval
ProverbEval: Benchmark for Evaluating LLMs on Low-Resource Proverbs
This dataset accompanies the paper:"ProverbEval: Exploring LLM Evaluation Challenges for Low-resource Language Understanding"ArXiv:2411.05049v3
Dataset Summary
ProverbEval is a culturally grounded evaluation benchmark designed to assess the language understanding abilities of large language models (LLMs) in low-resource settings. It consists of tasks based on proverbs in five languages:
Amharic
Afaan… See the full description on the dataset page: https://huggingface.co/datasets/israel/ProverbEval.AfriGuard
AfriGuard: Safety Evaluation Data for African Languages
AfriGuard is a human-annotated safety dataset covering 10 African languages: Amharic, Hausa, Igbo, Oromo, Shona, Swahili, Twi, Wolof, Yoruba, and Zulu. Each example contains a culturally grounded prompt/response pair in English and the target language, labeled with a safety top category, a safe/unsafe label, and majority-vote annotations from three native-speaker annotators.
Splits
Each language config… See the full description on the dataset page: https://huggingface.co/datasets/israel/AfriGuard.ur3-bimanual-lingbot-va
UR3 Bimanual Robot Dataset for LingBot-VA Fine-tuning
202 teleoperated episodes of a bimanual UR3 robot performing manipulation tasks, preprocessed and ready for LingBot-VA fine-tuning.
Dataset Summary
Property
Value
Episodes
202
Unique tasks
97
Total frames
61,074 (at 30 fps)
Cameras
3 (top, left wrist, right wrist)
Action space
30-dim (14 active: both arms EEF + grippers)
Format
LeRobot v2.1 + LingBot-VA pre-extracted latents
Task… See the full description on the dataset page: https://huggingface.co/datasets/ISRHUMANOID/ur3-bimanual-lingbot-va.frontend_dpo
DPO JavaScript Dataset
This repository contains a modified and expanded version of a closed-source JavaScript dataset. The dataset has been adapted to fit the DPO (Dynamic Programming Object) format, making it compatible with the LLaMA-Factory project. The dataset includes a variety of JavaScript code snippets with optimizations and best practices, generated using closed-source tools and expanded by me.
License
This dataset is licensed under the Apache 2.0 License.… See the full description on the dataset page: https://huggingface.co/datasets/israellaguan/frontend_dpo.isr-aloha-transfer-cube-experiment
Does ISR (Information-Standardized Trajectory Resampling) improve policy success? A closed-loop test on public human teleop data
TL;DR — On lerobot/aloha_sim_transfer_cube_human (50 human demos) with ACT, evaluated by 100 sim rollouts per cell over 2 training seeds:
claim
result
ISR beats the paper's 3× uniform baseline at equal frame budget (~⅓ frames)
42 / 40 % vs 31 / 20 % (seed 1 / seed 2) ✅
ISR at half the frames matches full data
60 / 62 % vs 62 / 63 % ✅… See the full description on the dataset page: https://huggingface.co/datasets/Kavin60606/isr-aloha-transfer-cube-experiment.kwaiklear-sample-level-agent-trajectories-2.2M