berth / README.md
JayPatel2574's picture
Dataset card: same layout on Kaggle
9a8e69e verified
|
Raw History Blame Contribute Delete
5.61 kB
metadata
license: apache-2.0
pretty_name: Berth
language:
  - en
tags:
  - code
  - software-engineering
  - agents
  - swe-bench
  - gemma
size_categories:
  - n<1K
task_categories:
  - text-generation
configs:
  - config_name: default
    data_files:
      - split: core
        path: core.jsonl
      - split: extended
        path: extended.jsonl
      - split: overlap
        path: overlap.jsonl

Berth: harness-compatible, post-cutoff Python tasks for offline Gemma 4 coding agents

Berth is a set of 653 verified bug-fix and feature tasks from 24 permissively licensed Python libraries. Every task is packaged in the data layout of the Gemma 4 Developer Agent Competition harness and installs and tests with no network. The layout covers tasks.jsonl columns, history-free repository snapshots, and code graphs in the official format.

It is meant as held-out validation and as a training pool for agents built for that harness. The competition's public development set has 129 tasks from 4 repositories.

Split Tasks Pass the official grader Statement Use
core 301 281 real GitHub issue evaluation
extended 350 297 pull-request description training
overlap 2 2 real issue also in SWE-rebench; kept apart

Recommended headline subset: core tasks with harness_compatible == true and post_cutoff == true (209 tasks). In these, both the issue and the pull request postdate Gemma 4's January 2025 pretraining cutoff, and the task passes the competition's own grading rules.

How every task was checked

  1. Fail before, pass after: in a fresh Python 3.13 environment, tests touched by the fix fail without it and pass with it, identically in two runs.
  2. Offline: inside a network-less namespace, the environment installs only from wheels/; the empty patch fails FAIL_TO_PASS and the gold patch passes.
  3. Stable: three more gold replays through an agent runner's tools and grader, under parallel load.
  4. Official grader (harness_compatible): the harness's own Phase 2 sequence is reproduced offline, with setup_workspace_test_config and _validate_junit_xml imported from the competition's swegemma wheel. The whole target test files must pass with -p no:anyio, and required tests are matched by name prefix. The gold patch must resolve and the empty patch must not.

Tasks failing step 4 are kept and flagged. Most use async test suites, which the official grader cannot run. To satisfy the official validator's prefix matching, some test ids were removed; they are listed per task in FAIL_TO_PASS_dropped_validator and PASS_TO_PASS_dropped_validator, and every task keeps at least one FAIL_TO_PASS test. For 33 tasks, dependencies were resolved as of the fix commit's date (env_pinned_to_commit_date).

Files

Path Content
tasks.jsonl (+ core/extended/overlap.jsonl) One task per line (fields below)
snapshots/<instance_id>.tgz Repository at base_commit as a fresh one-commit git repo (no forward history)
graphs/<repo>_<commit>.json Code graph, official node-link format (built with an Apache-2.0 builder that reproduces all 127 official graphs)
embeddings/<repo>_<commit>.npz 256-d stand-in vectors (deterministic random; the official encoder is unpublished)
wheels/ Every wheel any task needs, CPython 3.13 / manylinux x86_64
runs/gemma-4-31b-it-qat-w4a16-ct/ Reference-agent trajectories (.traj.jsonl), our grading (.json), official grading (.official.json)
results/ Localization baseline per task; every number in the paper (paper_numbers.json)

Task fields

Official columns: instance_id, repo, base_commit, patch, test_patch, problem_statement, hints_text, created_at, FAIL_TO_PASS, PASS_TO_PASS.

Berth fields:

  • split, statement_source (issue / issue_ref / pr), issue_urls, pr_url, fix_commit, merged_at, issue_created_at
  • post_cutoff: issue and PR both created on or after 2025-02-01
  • harness_compatible, env_pinned_to_commit_date
  • PASS_TO_PASS_flaky, FAIL_TO_PASS_dropped_validator, PASS_TO_PASS_dropped_validator
  • localization: files, nodes_edited and nodes_added (official graph node ids), module_level, aux_files
  • quality: underspecified (the hidden tests call a name the fix introduces and the issue never mentions), new_api_in_tests, f2p_count, patch_lines, files_changed, statement_chars
  • requirements, build_requires, graph, embeddings, snapshot, test_patch_files, patch_files

Baselines (core)

  • Localization (292 tasks whose fix edits a function or class). BM25 puts a gold function in its top 10 for 64.0% of tasks, against 9.9% for random. Graph propagation trades top-1 function accuracy (−6.9 points) for top-1 file accuracy (+6.5).
  • Gemma 4 31B (gemma-4-31b-it-qat-w4a16-ct on Kaggle 4×L4, simple single agent, 20-minute budget, official grading): 19.6% of the 281 compatible tasks, and 22.0% of the 209 post-cutoff ones. In 59% of failures the patch touched none of the files the real fix changes.

Licence

Each snapshot keeps the licence of its source repository (MIT, BSD or Apache-2.0; see LICENSES/). Berth's annotations, results and pipeline code are released under Apache-2.0. Pipeline code: https://github.com/JayPatel2574/berth

Citation

Jay Patel. Berth: Harness-Compatible, Post-Cutoff Python Tasks for Offline Gemma 4 Coding Agents. Gemma 4 Developer Agent Paper Track, Kaggle, 2026.