Datasets:

akseljoonas
/

mimo-openenv-software

cwd stringclasses 2 values	dataset_type stringclasses 1 value	docker_image stringclasses 8 values	instance_id stringclasses 8 values	problem_statement stringclasses 8 values	test_command stringclasses 2 values	test_patch stringclasses 8 values	verifier_timeout_sec int64 1.8k 1.8k
/testbed	opensource-code	format-code-task-000636:latest	format-code-task-000636	SSLTransport error: unexpected keyword argument 'ca_certs' Using amqp version 5.0.1, I discovered an SSLTransport error. When I use amqp version 5.0.0b1, SSL works just fine. Here's the error I get when using amqp==5.0.1: `/usr/local/lib/python3.8/dist-packages/kombu/connection.py:283: in channel chan = self...	bash /testbed/mimo_test_command.sh	diff --git a/t/unit/test_transport.py b/t/unit/test_transport.py index d94a520..a6a1a4e 100644 --- a/t/unit/test_transport.py +++ b/t/unit/test_transport.py @@ -700,6 +700,74 @@ class test_SSLTransport: with pytest.raises(socket.timeout): self.t._read(64) + def test_wrap_socket_sni_accepts_ca...	1,800
/workspace/repo	opensource-code	format-code-task-000666:latest	format-code-task-000666	I want `pyromat.get(idstr: str) -> object` so application code can retrieve one already-loaded substance from the PYroMat catalog by its exact substance ID. For example, after `import pyromat as pm`, `pm.get("ig.O2")` should return the oxygen substance object whose `sid()` is `"ig.O2"` and whose `pmclass()` is `"ig2"`;...	bash /workspace/repo/mimo_test_command.sh	diff --git a/usercase-test-coderl/test_get_catalog_lookup.py b/usercase-test-coderl/test_get_catalog_lookup.py new file mode 100644 --- /dev/null +++ b/usercase-test-coderl/test_get_catalog_lookup.py @@ -0,0 +1,177 @@ + + + + + + +import builtins +import copy +import json +import os +import socket +import subprocess +i...	1,800
/testbed	opensource-code	format-code-task-000932:latest	format-code-task-000932	Enforce schema when using json_writer() or json_reader() Hi, I have reproduced the example in https://fastavro.readthedocs.io/en/latest/json_writer.html but I've changed some items of the `records` array in order to introduce some type mismatches. ``` from fastavro import json_writer, parse_schema from io impor...	bash /testbed/mimo_test_command.sh	diff --git a/test_commands.json b/test_commands.json new file mode 100644 index 0000000..5c87055 --- /dev/null +++ b/test_commands.json @@ -0,0 +1,3 @@ +{ + "test_commands": ["python -m pytest tests/test_json.py -k \"json_writer_validator or json_writer_custom or json_writer_no_validator or json_writer_issue\" -v"] +}...	1,800
/testbed	opensource-code	format-code-task-001457:latest	format-code-task-001457	[FEAT] Make `secrets` parameter optional so we can use this action just to get the token. Is your feature request related to a problem? Please describe. The `secrets` parameter should not be mandatory. There is the option for `exportToken`, which is what I want to use to configure the Vault Terraform provider. How...	bash /testbed/mimo_test_command.sh	diff --git a/src/action.test.js b/src/action.test.js index 1837eaf..9c585d8 100644 --- a/src/action.test.js +++ b/src/action.test.js @@ -332,3 +332,85 @@ with blank lines expect(core.setOutput).toBeCalledWith('key', multiLineString); }) }); + +describe('parseSecretsInput optional', () => { + it('retur...	1,800
/testbed	opensource-code	format-code-task-001547:latest	format-code-task-001547	Support react-i18next "defaults" prop #### Version - i18next: <i>11.2.2</i> - i18next-scanner: <i>2.4.6</i> - react-i18next: <i>7.7.0</i> #### Configuration ```js options: { // use strings as keys nsSeparator: false, keySeparator: false, // settings defaultNs: "frontend", lngs: ["en"], resource: {...	bash /testbed/mimo_test_command.sh	diff --git a/test/parser.js b/test/parser.js index 959c46d..9c532ae 100644 --- a/test/parser.js +++ b/test/parser.js @@ -198,6 +198,81 @@ test('Parse Trans component #3', (t) => { t.end(); }); +test('Parse Trans component with defaults attribute on self-closing tag', (t) => { + const parser = new Parser({ + ...	1,800
/workspace/repo	opensource-code	format-code-task-001661:latest	format-code-task-001661	I want the friendship model API to treat a stored friendship row as a bidirectional relationship between two Django users. The library should expose `Friendship.objects.friends_for_user(user)`, returning a list of dictionaries where each item has `friend` set to the other `User` and `friendship` set to the matching `Fr...	bash /workspace/repo/mimo_test_command.sh	diff --git a/usercase-test-coderl/test_bidirectional_friendship.py b/usercase-test-coderl/test_bidirectional_friendship.py new file mode 100644 --- /dev/null +++ b/usercase-test-coderl/test_bidirectional_friendship.py @@ -0,0 +1,257 @@ + + + + + + + + + + + + + + +import pytest + +from django.conf import settings + + +...	1,800
/workspace/repo	opensource-code	format-code-task-002105:latest	format-code-task-002105	I want the Python client to support creating custom Ollama models and uploading model file blobs through the public client API. A synchronous user should be able to call `Client.create(model: str, quantize: str \| None = None, from_: str \| None = None, files: dict[str, str] \| None = None, adapters: dict[str, str] \| None...	bash /workspace/repo/mimo_test_command.sh	diff --git a/usercase-test-coderl/test_custom_model_creation.py b/usercase-test-coderl/test_custom_model_creation.py new file mode 100644 --- /dev/null +++ b/usercase-test-coderl/test_custom_model_creation.py @@ -0,0 +1,503 @@ +import importlib +import json +import threading +from http.server import BaseHTTPRequestHand...	1,800
/testbed	opensource-code	format-code-task-002658:latest	format-code-task-002658	`extendEnv` does not work when combined with `shell: true` The `extendEnv` option does not work as intended when combined with the `shell: true` option: ```js process.env.TEST = 'test' const result = execa.sync('echo $TEST', { shell: true, env: {}, extendEnv: true }) assert.equal(result.stdout, 'test') ``` `s...	bash /testbed/mimo_test_command.sh	diff --git a/test.js b/test.js index 91fd135..ab49352 100644 --- a/test.js +++ b/test.js @@ -487,6 +487,31 @@ test('do not extend environment with `extendEnv` option', async t => { ]); }); +test('use extendEnv with shell:true and explicit empty env (sync)', t => { + const {stdout} = m.sync('echo $FOO', {shell: tru...	1,800

MiMo software engineering for OpenEnv

An OpenEnv adapter for eight software engineering tasks from XiaomiMiMo/MiMo-V2.6-RL-oss. Each task retains its original problem statement, repository image, test patch, test command and binary reward: 1 if the original tests exit successfully, 0 otherwise. Infrastructure errors are reported as errors. No LLM judge or external service credentials are used.

The tasks cover Python and JavaScript library repairs. tasks.jsonl lists task identities and immutable upstream images; tasks/ contains the original eight task records. Selection favored smaller images and self-contained tests. This is a small, non-random subset, not a reproduction of MiMo training or an estimate of general software engineering ability.

Source: https://github.com/akseljoonas/mimo-openenv-software

Dataset: https://huggingface.co/datasets/akseljoonas/mimo-openenv-software

Runtime contract

Each task has a separate public linux/amd64 image tagged ghcr.io/akseljoonas/mimo-openenv-software:TASK_ID-v2. The image starts its own OpenEnv server on port 8000 without mounted files, secrets or additional environment variables. Use an OpenEnv WebSocket session at /ws: reset accepts task_id; step accepts {"operation":"shell","command":"..."} or {"operation":"finish"}. Standalone HTTP /reset and /step are stateless in the pinned OpenEnv release.

Reset restores the repository snapshot. Shell actions start in the task's original working directory, run as UID 2000, last at most 60 seconds, and return at most 12,000 output bytes. finish invokes Xiaomi's original verifier and returns done: true; reaching 64 actions also grades. A terminal session returns its cached verdict until reset. Tests are installed only while grading. All runtime verifier commands and patch writes also run as UID 2000; the server and stored task record are inaccessible to that user. This is not a security boundary against every possible reward exploit in arbitrary upstream repositories.

Build preparation uses Xiaomi's existing Git-history stripping routine, then its setup assertion, before snapshotting the repository. This removes reachable future fixes present in some original images. Existing task files and grading code are not rewritten. The adapter has its own Python runtime so task dependencies remain intact.

Pinned sources

Dataset: 639865fd3374018d6cb29b9fb82dd531406fcf5f (Apache-2.0).
MiMo-Agent: 467f0a19016f0ac4d63b8d17a1f0da9ba07f232c (MIT).
OpenEnv: 86a180ede21e044f7929b9a7783ad83aa67d83a3, Arena's required revision.
Python runtime base and task images are pinned by digest. Python packages are pinned in requirements.lock.

The original code-task runner is Xiaomi's recipes/code/mimoagent_runner.py at verl revision a2ad9f6. It constructs the dataset environment, runs an agent, and calls calculate_reward. This adapter replaces that agent transport with OpenEnv while reusing the published OpenSourceCodeEnvironment grading implementation. XiaomiMiMo owns the upstream task/verifier; this repository owns only the OpenEnv adapter and selection. Upstream project licenses remain applicable inside each image.

Build and verify

Install the pinned OpenEnv CLI. For a row in tasks.jsonl:

export DOCKER_DEFAULT_PLATFORM=linux/amd64
openenv build -t ghcr.io/akseljoonas/mimo-openenv-software:TASK_ID-v2 \
  --build-arg TASK_ID=TASK_ID --build-arg TASK_IMAGE=SOURCE_IMAGE_DIGEST
python verify.py IMAGE TASK_ID --output evidence/TASK_ID.json

verify.py runs the image with 2 CPUs and 16 GiB, checks readiness and all six OpenEnv endpoint checks, then replays real WebSocket episodes: unsolved reward 0, stable terminal reward, protected verifier, absent future Git refs, and reset isolation. The Django task additionally uses --solution solutions/friendship.py to check reward 1 and restoration to reward 0. Reference solutions are excluded from images. Reports document the checks actually performed; they do not establish Arena admission, training or private evaluation results.

The Publish images GitHub workflow builds, tests and publishes the selected images with short-lived GitHub Actions registry authentication. No Hugging Face token is included in this repository or its artifacts. Submissions are separate from the build workflow; the account owner has authorized the ongoing experiment series within Arena's rolling daily quota.

The prepared Arena request allows 180 seconds for reset, 1,200 seconds for rollout and 120 seconds for verification, with 64 tool calls, a 32,768-token context and a 10 GiB workspace (Arena’s documented default). These Arena budgets constrain the unchanged upstream verifier, whose internal timeout remains 1,800 seconds.

Status

All eight v2 images passed native Linux CI, anonymous pulls and real WebSocket episode replays. All tasks terminate at reward 0 when unsolved; the Django reference repair scores 1 and resets back to 0. The tests also cover reset isolation, stable terminal rewards, identical schemas and unprivileged verifier Git hooks. The largest image is 674 MB compressed; shared layers total 4.73 GB.

The exact accepted, digest-pinned request is in submission.json; evidence/preflight.json summarizes validation and its limits. The first request was refused before acceptance because its 44 GiB workspace plus image and headroom exceeded the sandbox disk limit; no slot was consumed. The corrected request uses 10 GiB, after measuring writable layers of at most 180 MB following reset and grading across all eight tasks. It was accepted on October 8, 2026, and passed all 38 real admission checks. V1 images were superseded by the verifier privilege fix and should not be submitted.

Baseline run: 03bfc81de59cf3e53b9e05d0 on Trackio. Its pinned dataset revision is 4ce4fe143d71ca6365bd187e96efa15bd7ab2787. GPU training was observed at its first optimizer update; no private evaluation result has yet been observed. Training metrics and private scores are tracked separately.

Five prepared request variants and their hypotheses are recorded in experiments/queue.json. They reuse the same validated images and original graders. Prepared requests have not been submitted. See experiments/README.md for selection and monitoring rules.

Downloads last month: -

Number of rows:

Total file size:

423 kB