SWE-Gym + Claude Code Quickstart

中文版 · Full experiment guide

This quickstart prepares the public Dressage SWE-Gym recipe, runs Claude Code inside E2B task templates, evaluates every patch in a fresh sandbox, and starts the synchronous Qwen3.5-4B GRPO recipe.

1. Install the repository and data dependencies

git submodule update --init --recursive

python3 -m pip install -e .
python3 -m pip install --no-build-isolation blackbox_server/
python3 -m pip install pyarrow huggingface_hub
python3 -m pip install \
  'swegym @ git+https://github.com/SWE-Gym/SWE-Bench-Package.git@16dd480cce9b27bf111a362d280881c6def5d2a7'

The pinned SWE-Gym package is used only while preparing data. The generated evaluation command embeds the official repository-specific test script and log parser, so task sandboxes do not need the swegym Python package.

2. Prepare E2B templates

Each SWE-Gym row needs a fresh task image containing:

• the repository and its original test environment under /testbed;
• Claude Code;
• this repository's Blackbox Server;
• Blackbox Server listening on port 31000.

Build a wheel from the public Blackbox Server and package an isolated runtime:

python3 -m pip install build uv
python3 -m build --wheel --outdir dist blackbox_server/

export CLAUDE_CODE_BBS_VERSION=1.1.0
export CLAUDE_CODE_BBS_WHEEL_URL="$PWD/dist/dressage_blackbox_server-1.1.0-py3-none-any.whl"
export CLAUDE_CODE_ARTIFACT_DIR="$PWD/data/claude-code-artifacts"

bash dressage/recipes/swegym/prepare_claude_code_sandbox_artifacts.sh

Build one E2B template per distinct SWE-Gym task image and write a JSON object mapping each Docker image to its E2B template name:

{
  "xingyaoww/sweb.eval.x86_64.example:latest": "e2b-swegym-example"
}

Use the repository utility to enumerate images, build each template, and smoke-test it:

python3 examples/data/swegym/prepare_swegym_e2b.py list-images \
  --download \
  --download-dir data/swegym-source \
  --split train \
  --output data/swegym-images.txt

export E2B_API_KEY=e2b_...
export TASK_IMAGE=xingyaoww/sweb.eval.x86_64.example:latest
export TEMPLATE_NAME=e2b-swegym-example

python3 examples/data/swegym/prepare_swegym_e2b.py build
python3 examples/data/swegym/prepare_swegym_e2b.py smoke

Repeat the build for every image in data/swegym-images.txt, assigning each a unique template name, and record the resulting mapping. See the full experiment guide for details.

3. Convert SWE-Gym data

Convert the downloaded 293-row training split:

python3 examples/data/swegym/prepare_swegym_data.py \
  data/swegym-train-claude-code-e2b.jsonl \
  --input data/swegym-source/train.parquet \
  --split train \
  --provider e2b \
  --sandbox-image-map data/e2b-template-map.json \
  --blackbox-type claude_code \
  --max-turns 80 \
  --permission-mode acceptEdits

For a one-row conversion smoke test, add --limit 1. The converter verifies the complete split before applying the limit. The final JSONL contains no gold patch. It includes:

• a mandatory before-agent Git sanitizer;
• the fixed official SWE-Gym evaluator and log parser;
• task-specific FAIL_TO_PASS and PASS_TO_PASS tests;
• Claude Code backend options, including working_directory=/testbed;
• the E2B template mapped for each task.

4. Smoke-test a template

Before allocating training GPUs, verify that a template resumes Blackbox Server and exposes port 31000:

python3 examples/data/swegym/prepare_swegym_e2b.py smoke \
  --template-name e2b-swegym-example

Also verify one real task end to end: Claude Code must produce a patch under /testbed, and the fresh evaluation sandbox must emit a DRESSAGE_SWEGYM_REWARD_JSON= marker.

5. Start synchronous GRPO

Prepare Qwen3.5-4B Hugging Face and Megatron distributed checkpoints under the same model root, then run:

export MODEL_ROOT=/path/to/models
export PROMPT_DATA="$PWD/data/swegym-train-claude-code-e2b.jsonl"
export DRESSAGE_SANDBOX_PROVIDER=e2b
export DRESSAGE_E2B_API_KEY=e2b_...
export DRESSAGE_E2B_BLACKBOX_PORT=31000
export DRESSAGE_PROXY_URL=https://proxy.example.com

bash examples/scripts/run_dressage_swegym_qwen3.5_4b_claude_code_sync_4_node.sh

DRESSAGE_PROXY_URL must be an HTTP(S) endpoint that the E2B sandboxes can reach.

The launcher explicitly selects dressage.rollout.generate.blackbox_dispatch_swegym.generate. The generic blackbox dispatcher remains recipe-agnostic; the SWE-Gym dispatcher owns fresh evaluation and trajectory-integrity checks.

Reference defaults reproduce the reviewed experiment shape: TP2/CP4, 8 prompts × 16 samples, global batch size 128, 500 rollout updates, normalized GRPO advantages, vanilla token-level TIS, and low-variance KL loss coefficient 0.001. Override infrastructure and topology through environment variables.