README.md

metadata

license: other
pretty_name: FARBench Docker Images
tags:
  - benchmark
  - docker-image
  - research-agent

FARBench Docker Images

Per-task Docker images, saved with docker save | gzip and uploaded as plain LFS files. Each tarball is a fully self-contained image (CUDA + Python + task deps + baked task data); load it with docker load and the resulting image carries the canonical tag farbench/farbench:<task>-<cuda>.

Files are at the repo root with the flat naming convention <task>-<cuda>.docker.tar.gz (e.g. mnist_classification-cu118.docker.tar.gz).

Quick start

# 1. Download the tarball you need (single task, cu118 example).
huggingface-cli download \
    FARBenchAnonymous/FARBench \
    mnist_classification-cu118.docker.tar.gz \
    --repo-type dataset \
    --local-dir ./farbench-images

# 2. Load into your Docker daemon.
docker load -i ./farbench-images/mnist_classification-cu118.docker.tar.gz

# 3. The image is now available locally as e.g.
#    farbench/farbench:mnist_classification-cu118
docker images | grep farbench

If a tarball is split (*.part00, *.part01, ...) merge first:

cat <task>-<cuda>.docker.tar.gz.part* > <task>-<cuda>.docker.tar.gz
docker load -i <task>-<cuda>.docker.tar.gz

Available tasks

Task	Domain	Metric	cu118	cu128
ade20k	computer vision	mIoU	ade20k-cu118	ade20k-cu128
aime_math_rl	natural language processing	exact_match	aime_math_rl-cu118	aime_math_rl-cu128
assist2009_kt	natural language processing	auc_roc	assist2009_kt-cu118	assist2009_kt-cu128
asvspoof2021_la	audio/speech understanding	eer	asvspoof2021_la-cu118	asvspoof2021_la-cu128
bigcodebench_codegen	natural language processing	pass_at_1	bigcodebench_codegen-cu118	bigcodebench_codegen-cu128
cifar100lt	computer vision	balanced_accuracy	cifar100lt-cu118	cifar100lt-cu128
cifar100n	computer vision	accuracy	cifar100n-cu118	cifar100n-cu128
climsim_lowres	AI for science	mean_r2	climsim_lowres-cu118	climsim_lowres-cu128
clotho_caption	Audio/Speech	spider	clotho_caption-cu118	clotho_caption-cu128
cogniplan	robotics	exploration_score	cogniplan-cu118	cogniplan-cu128
crohme_hmer	computer vision	exprate	crohme_hmer-cu118	crohme_hmer-cu128
div2k_sr_x4	Computer Vision	psnr_y	div2k_sr_x4-cu118	div2k_sr_x4-cu128
domainnet_quickdraw	computer vision	accuracy	domainnet_quickdraw-cu118	domainnet_quickdraw-cu128
etth1_forecasting	AI for science	mse	etth1_forecasting-cu118	etth1_forecasting-cu128
flip_aav	AI for science	spearman_rho	flip_aav-cu118	flip_aav-cu128
habitat3	robotics	nav_seek_success	habitat3-cu118	habitat3-cu128
humanoidbench	robotics	success_rate	humanoidbench-cu118	humanoidbench-cu128
iwildcam_wilds	computer vision	macro_f1	iwildcam_wilds-cu118	iwildcam_wilds-cu128
ljspeech_tts	Audio/Speech	utmos	ljspeech_tts-cu118	ljspeech_tts-cu128
metrla_traffic	AI for science	mae_60min	metrla_traffic-cu118	metrla_traffic-cu128
minigrid	robotics	success_rate	minigrid-cu118	minigrid-cu128
mnist_classification	Computer Vision	accuracy	mnist_classification-cu118	mnist_classification-cu128
objaverse_3dgen	computer vision	lpips	objaverse_3dgen-cu118	objaverse_3dgen-cu128
ogbg_molpcba	AI for science	avg_precision	ogbg_molpcba-cu118	ogbg_molpcba-cu128
qlib_stock	natural language processing	ic_mean	qlib_stock-cu118	qlib_stock-cu128
qm9	AI for science	mae	qm9-cu118	qm9-cu128
scanobjectnn	Computer Vision	overall_accuracy	scanobjectnn-cu118	scanobjectnn-cu128
screenspot_pro	computer vision	grounding_score	screenspot_pro-cu118	screenspot_pro-cu128
split_cifar100	computer vision	average_accuracy	split_cifar100-cu118	split_cifar100-cu128
terra_incognita	computer vision	balanced_accuracy	terra_incognita-cu118	terra_incognita-cu128
vlabench_manipulation	robotics	success_rate	vlabench_manipulation-cu118	vlabench_manipulation-cu128
voicebank_demand	audio/speech understanding	pesq	voicebank_demand-cu118	voicebank_demand-cu128
weatherbench_z500t850	AI for science	rmse_z500	weatherbench_z500t850-cu118	weatherbench_z500t850-cu128
wilds_fmow	computer vision	worst_region_accuracy	wilds_fmow-cu118	wilds_fmow-cu128

CUDA variants

• *-cu118.docker.tar.gz — built on nvidia/cuda:11.8.0-runtime-ubuntu22.04.
• *-cu128.docker.tar.gz — built on nvidia/cuda:12.8.1-runtime-ubuntu22.04 (e.g. RTX 5090).

Use cu118 unless your GPU requires CUDA 12.x kernels. Both variants produce identical task behaviour and are interchangeable from the agent's point of view.

License

The RABench framework is released under Apache-2.0. The bundled datasets and pre-cached model weights are redistributed from their original sources and retain the original licenses; see each task's README in the data repository.