README.md

metadata

license: apache-2.0
tags:
  - docker
  - alfworld

e2l-train Docker image

Prebuilt Stage-2 training image for Ch1nyzzz/e2-learning.

Code and environment are both inside the image. You do not need to git clone to train.

Contents:

• /workspace/e2-learning: this repository (scripts, configs, .env.example)
• /opt/verl-agent: langfengQ/verl-agent at 20bd331 plus the stage-2 patch
• env A (system Python 3.12): torch 2.8.0 / vllm 0.11.0 / flash-attn 2.8.3
• env B (/opt/venvs/rwml, Python 3.10): torch 2.6.0 / vllm 0.8.3 / verl 0.4.1

Tag after load: e2l-train:latest (~46.6GB uncompressed, ~21GB this tarball). Working directory: /workspace/e2-learning.

Load

hf download erv1n/e2l-train-image --repo-type dataset --local-dir ./e2l-train-image
cd e2l-train-image
sha256sum -c e2l-train.tar.gz.sha256
gunzip -c e2l-train.tar.gz | docker load

Then follow the GitHub playbook (source of truth, may be newer than the copy inside the image): https://github.com/Ch1nyzzz/e2-learning/blob/main/docs/playbook.md

Create a host work directory, docker run with those mounts, pull models, start training. Do not bind-mount a git checkout over /workspace/e2-learning or you will hide the baked-in code.