{"data":{"kind":"file","path":"README.md","version_id":"cadrbcidonlwfutz66xixyp4","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":3517,"modified_at":"2026-09-13T04:06:06.954000","content_hash":"22bc419a5c8f19c3d4e5e7f21eae09e54e1d914ce1669ff1c98eb452c56b11ca"},"entries":[],"content":"# multiswe\n\nMulti-SWE-bench issue-resolving tasks across many languages, each backed by a prebuilt sandbox image. Scored by extracting a source-only fix patch, running the task's `fix-run.sh` harness, and validating the resulting report against the expected fail/pass tests via upstream `multi_swe_bench` report generation.\n\n## Taskset\n\n- **Source:** [`PrimeIntellect/Multi-SWE-RL-Verified`](https://huggingface.co/datasets/PrimeIntellect/Multi-SWE-RL-Verified)\n- **Size:** 2,232 tasks (`train` split)\n\n## Changelog\n\n- 2026-09-04: Add a solver system prompt (adapted from FrontierCode 1.1's fair-internet-use prompt) that allows documentation lookups but forbids retrieving the task's upstream fix online or from non-current-branch git history; complements #795, which restored network access.\n- 2026-09-03: Restore default solver network access by removing the `network_allow=[]` override introduced in #780; training rollouts need outbound network.\n- 2026-08-31: Yield task records on demand so bounded evaluations construct only the requested prefix.\n- 2026-08-27: Install `multi-swe-bench==1.1.2` from its PyPI source archive because the published wheel contains conflicting duplicate entries and current `uv` rejects it; restore the lowercase Qiskit adapter omitted by that source archive.\n- 2026-07-17: `patch_capture` now imports from `verifiers.v1` (`capture_patch`/`resolve_head` upstreamed in verifiers#2054); the copied module is removed. Requires `verifiers>=0.2.2.dev5`.\n- 2026-07-17: Task images now use the raw source refs (`docker_image` as-is, or the upstream Docker Hub `mswebench/<org>_m_<repo>:pr-<number>` ref) instead of rewriting to `prime/primeintellect/`. The Prime platform serves these natively (org-less platform copies under the source names; missing ones auto-import from Docker Hub on first use), so no registry mapping is needed.\n- 2026-07-16: `finalize` persists the agent's final patch to `trace.info[\"patch\"]` (truncated at 2 MB with `patch_truncated`; failures record `patch_error` instead of failing the rollout) via the copied `patch_capture` helper.\n- 2026-07-10: Ported to the task-centric verifiers API: rewards and lifecycle hooks live on the `Task` (a `TaskData` row + behavior split), and task-facing config knobs (judges, tool/user placement, scoring parameters) moved from `--env.taskset.*` to `--env.taskset.task.*`. Requires `verifiers>=0.2.0` and Python `>=3.11`.\n- 2026-07-08: Task images now resolve from the public Prime platform registry (`prime/primeintellect/<name>:<tag>`, `mswebench/` namespace stripped) instead of the private GCP Artifact Registry mirror; any Prime user can pull them.\n- 2026-07-06: Default dataset renamed to `PrimeIntellect/Multi-SWE-RL-Verified` (formerly `Multi-SWE-RL`); the old name redirects.\n- 2026-07-01: Added `filter_fn`, applied directly with `datasets.Dataset.filter` to raw HF rows before task construction.\n- 2026-07-01: Default dataset is now `PrimeIntellect/Multi-SWE-RL`. Known PrimeIntellect datasets are typed on `MultiSWEConfig.dataset_name`; known datasets get split validation for their available splits.\n- 2026-06-30: Setup now moves top-level `/home` grading scripts and `/home/test.patch` through a host-side archive and removes them from the sandbox. Scoring restores the archived artifacts before running `/home/fix-run.sh`. This intentionally diverges from the original Multi-SWE images, where those artifacts remain visible, because the v1 live sandbox would otherwise let agents read or edit grading material.\n","encoding":"utf-8","truncated":false,"total_bytes":3517},"status":null}