harbor-rl-data

Harbor task datasets for online RL training with rllm-harbor.

Each task is a self-contained Harbor task directory containing:

• task.toml — task metadata (resources, timeouts)
• instruction.md — problem statement shown to the agent
• tests/test.sh — evaluation script (runs agent patch, grades with swebench)
• tests/config.json — install config (python version, packages, test command)

Reward is 1 if all FAIL_TO_PASS tests pass and all PASS_TO_PASS tests remain passing, else 0.

---

Dataset versions

v0.0.2 — v0.0.2/ (latest)

Split	Tasks	Source parquet
train	877	swerebench_sc4to7_877instances.parquet
val	500	harbor_swe_val_tasks

Changes vs v0.0.1:

• SWE-bench fork upgraded to 4.1.0 with install_config bypass (supports 877 custom repos beyond the 66 in upstream MAP_REPO_VERSION_TO_SPECS)
• Fixed reward=0 bug caused by modal_eval → pydantic import chain in containers
• Grader inlined as PARSER_EOF heredoc in test.sh (no separate parser.py)
• Train task resources aligned with val: cpus=1, memory_mb=4096, storage_mb=10240
• Instruction template unified with official SWE-bench format
• Wheel: swebench-4.1.0-py3-none-any.whl (fork with install_config support)

---

v0.0.1 — v0.0.1/ (archived)

Split	Tasks	Source parquet
train	877	swerebench_sc4to7_877instances.parquet
val	500	harbor_swe_val_tasks

Initial release. Uses swebench-4.0.3 fork. Parser ran as a separate parser.py file.

---

Usage

from huggingface_hub import snapshot_download

# Download latest train tasks
snapshot_download(
    repo_id="ElvinDu518/harbor-rl-data",
    repo_type="dataset",
    local_dir="/data/harbor_swe_tasks",
    allow_patterns="v0.0.2/harbor_swe_tasks/**",
)

Or clone and point the Harbor task registry at the local directory.