KouShi2/ailab-code-evaluation-images/tb3 - Storage Bucket

TB3.0 Evaluation Images

Docker images for 74 TB3.0 evaluation tasks. Each task is distributed as two independent Docker archives:

• an agent image containing the task environment;
• a verifier image used to evaluate the task result.

The collection contains 148 .tar files in total. Agent and verifier archives are matched by their filename.

Directory layout

tb3/
├── README.md
└── images/
    ├── agent/
    │   ├── atrx-vep-crispr.tar
    │   └── ...
    └── verifier/
        ├── atrx-vep-crispr.tar
        └── ...

Requirements

• Docker
• Hugging Face CLI (pip install -U huggingface_hub)

Download one task

Replace atrx-vep-crispr with the desired task name:

hf buckets cp \
  hf://buckets/KouShi2/ailab-code-evaluation-images/tb3/images/agent/atrx-vep-crispr.tar \
  ./atrx-vep-crispr-agent.tar

hf buckets cp \
  hf://buckets/KouShi2/ailab-code-evaluation-images/tb3/images/verifier/atrx-vep-crispr.tar \
  ./atrx-vep-crispr-verifier.tar

Load the images

Each TAR is a standard archive produced by docker image save and contains exactly one image tag.

docker load -i ./atrx-vep-crispr-agent.tar
docker load -i ./atrx-vep-crispr-verifier.tar

The loaded tags follow this convention:

tb3-try2/<task-name>:agent-v3.0.0
tb3-try2/<task-name>:verifier-v3.0.0

For example:

tb3-try2/atrx-vep-crispr:agent-v3.0.0
tb3-try2/atrx-vep-crispr:verifier-v3.0.0

Download all images

hf cp \
  hf://buckets/KouShi2/ailab-code-evaluation-images/tb3/images/ \
  ./tb3-images/

This downloads the agent/ and verifier/ directories while preserving their layout. The full collection is large, so make sure sufficient disk space is available before downloading everything.

Verify the collection

After downloading the full collection, each directory should contain 74 TAR files:

find ./tb3-images/agent -maxdepth 1 -type f -name '*.tar' | wc -l
find ./tb3-images/verifier -maxdepth 1 -type f -name '*.tar' | wc -l

Expected output:

74
74

Use the agent and verifier archives with the same filename as a task pair.