MiMo-V2.6-RL-harbor-cyber

[](https://huggingface.co/spaces/HuggingFaceH4/harbor-visualiser?dataset=FineEnvs/MiMo-V2.6-RL-harbor-cyber) [](https://huggingface.co/spaces/FineEnvs/MiMo-RL-Envs-Explorer) # MiMo-V2.6-RL Cyber (Harbor) **Reproduce a real memory-safety crash (ARVO).** 1,000 Harbor tasks from the Cyber domain of Xiaomi's [MiMo-V2.6-RL-oss](https://huggingface.co/datasets/XiaomiMiMo/MiMo-V2.6-RL-oss), the RL environments MiMo-V2.6 was trained on, converted so every one runs as a standard Harbor task. The agent gets a sanitizer report and the project's source and fuzzer binary, and has to submit an input that crashes it in the expected function with the expected error. The image's own verify server re-runs the last submission: reward 1 for the right crash, 0 otherwise. | | | |---|---| | Tasks | 1,000 | | Graded by | PoC verify server (deterministic) | | Reference step limit | 300 | | Source | [`XiaomiMiMo/MiMo-V2.6-RL-oss`](https://huggingface.co/datasets/XiaomiMiMo/MiMo-V2.6-RL-oss) at `639865fd3374` | | Reference harness | XiaomiMiMo/verl `a2ad9f6` + XiaomiMiMo/mimoagent `467f0a1` | | Adapter version | `mimo_harbor 1.3.0` (see the changelog below) | Part of a set of six: [code](https://huggingface.co/datasets/FineEnvs/MiMo-V2.6-RL-harbor-code), [cyber](https://huggingface.co/datasets/FineEnvs/MiMo-V2.6-RL-harbor-cyber), [general](https://huggingface.co/datasets/FineEnvs/MiMo-V2.6-RL-harbor-general), [terminal](https://huggingface.co/datasets/FineEnvs/MiMo-V2.6-RL-harbor-terminal), [webdev](https://huggingface.co/datasets/FineEnvs/MiMo-V2.6-RL-harbor-webdev), [music](https://huggingface.co/datasets/FineEnvs/MiMo-V2.6-RL-harbor-music). ## Layout ```text tasks/<task_id>/ ├── task.toml # image pinned by digest, one-time setup (healthcheck), judge settings, provenance ├── instruction.md # the prompt the agent sees, word for word what Xiaomi's harness gives it ├── environment/ # Dockerfile (FROM the same digest) and setup/, a readable copy of the setup └── tests/ # test.sh and everything grading needs; uploaded only after the agent finishes registry.json # Harbor registry entry data/tasks.jsonl # one metadata row per task, for filtering without walking the tree manifest.json # sha256 of every task directory jobs/cyber.yaml # the reference agent settings agents/mimo_opencode.py ``` Example: [task.toml](https://huggingface.co/datasets/FineEnvs/MiMo-V2.6-RL-harbor-cyber/blob/main/tasks/arvo_10055/task.toml) · [instruction](https://huggingface.co/datasets/FineEnvs/MiMo-V2.6-RL-harbor-cyber/blob/main/tasks/arvo_10055/instruction.md) · [verifier](https://huggingface.co/datasets/FineEnvs/MiMo-V2.6-RL-harbor-cyber/blob/main/tasks/arvo_10055/tests/test.sh). ## Run ```bash hf download FineEnvs/MiMo-V2.6-RL-harbor-cyber --repo-type dataset --local-dir mimo-cyber cd mimo-cyber PYTHONPATH=agents HF_TOKEN=hf_... harbor run -y -c jobs/cyber.yaml # or on Daytona (builds the Dockerfile, runs the agent unprivileged natively — closer to Xiaomi's pods): DAYTONA_API_KEY=... DAYTONA_API_URL=... PYTHONPATH=agents HF_TOKEN=hf_... harbor run -y -c jobs/cyber.daytona.yaml ``` The job runs `agents/mimo_opencode.py` (OpenCode 1.18.32) on HF Sandbox with the reference settings: the step limit above, replies of up to 65,536 tokens, thinking low, no web fetch or search. Change `model_name` and the provider block for another model. Any Harbor agent can run these tasks; this one also applies two steps of Xiaomi's harness a task file can't express: the answer-leak blocklist after install, and the unprivileged `agent` user on sandboxes that run everything as root. Serving them to a trainer works like any Harbor dataset, for example with OpenEnv: ```bash openenv harbor serve --dataset FineEnvs/MiMo-V2.6-RL-harbor-cyber --llm-url http://127.0.0.1:8000/v1 --model <your-model> ``` ## How it was converted With the [mimo_harbor adapter](https://github.com/adithya-s-k/FineEnvs/tree/main/tooling/mimo-rl-explorer/mimo_harbor). It reads the source at one pinned revision and every image through a digest lock, so a rerun produces byte-identical tasks (`manifest.json`). Setup and grading follow Xiaomi's harness step by step, as the explorer's runner does, using the same vendored graders: - Nothing that grades a task (hidden tests, rubrics, verifier scripts) is reachable while the agent works. - A testbed failure (a patch that won't apply, a system that died, a judge that never answered) writes no reward, so Harbor records an error rather than a 0. - Images are pinned by digest; HF Sandbox and other prebuilt-image backends run them without a build step. ## Where this differs from Xiaomi's setup, in plain terms The task, its files, how it is graded, and the reward you get are **exactly Xiaomi's**. These are the only differences, and how we handle each: - **The agent is [OpenCode](https://opencode.ai), not Xiaomi's own agent.** The model gets the same instruction and the same reward — only the tool it uses to do the work is different. - **It can't look the answer up online.** Xiaomi ran with no internet. We let the model reach its own model API but block, by name, the places an answer could come from — code sites (GitHub, GitLab, …), search engines, this dataset's own page on the Hub, and package registries — so they don't resolve while it works. That's the whole list. The detailed, technical version is in the [adapter README](https://github.com/adithya-s-k/FineEnvs/tree/main/tooling/mimo-rl-explorer/mimo_harbor#deviations-from-the-reference-and-why). ## Validation - All 7,780 tasks across the six datasets load with Harbor's task loader and pass the adapter's static checks (digest-pinned images, setup payload equal to its readable copy, rubrics only under `tests/`). - With Harbor's no-op agent every dataset scores 0 and every verifier runs to completion: no free rewards. - **Parity with Xiaomi's harness.** Following Harbor's parity procedure, the 36-task parity subset (6 per dataset) was run 3 times on each side with the same agent (OpenCode 1.18.32), model (GLM-5.3 via deepinfra, thinking low), step limits, judges and sandboxes, in Harbor and in the [explorer](https://huggingface.co/spaces/FineEnvs/MiMo-RL-Envs-Explorer), which runs Xiaomi's harness. The score ranges overlap for all six datasets. For this one: | | Mean reward (mean ± SEM, 3 runs) | Runs | |---|---:|---| | Xiaomi's harness (explorer) | 0.822 ± 0.011 | 0.800, 0.833, 0.833 | | Harbor | 0.833 ± 0.096 | 1.000, 0.667, 0.833 | Full results, per-task outcomes and notes are in `parity.md` and `parity_experiment.json`. ## Patches, updates and fixes A timeline of changes to this dataset and the `mimo_harbor` adapter that builds it. Every task records the adapter version that produced it in `task.toml` (`[metadata].adapter`) and in `manifest.json`, so you can tell which of these a given copy includes. ### 1.3.0 — 2026-10-09 — environment-level answer-leak enforcement - The answer-leak blocklist is now applied to `/etc/hosts` as the **last step of task setup**, so the answer routes (GitHub and its API, PyPI / npm / crates, the Go module proxy, and the Hub's own dataset hosts) are blocked for **any** agent or trainer — not only the reference agent. Before this, the task only *staged* the list and the reference agent applied it after installing; a different agent on a network-connected sandbox could still reach those hosts and fetch a fix. - The reference agent (`agents/mimo_opencode.py`) lifts the blocklist to install OpenCode (fetched from GitHub), then re-applies it, so the model/judge router (`router.huggingface.co`) and the relay stay open throughout while the answer routes stay closed during the run. - Affects **code, cyber, general and terminal** (the domains that carry a blocklist); **webdev** and **music** are unchanged. - Prompted by independent third-party analysis of reward-hacking in the released MiMo environments. ### 1.2.0 — 2026-10-05 — initial public release - All 7,780 environments across six domains (code, cyber, general, terminal, webdev, music) published as Harbor tasks, set up and graded 1:1 with Xiaomi's harness (`verl` `a2ad9f6` + `mimoagent` `467f0a1`). - Anti-reward-hacking faithful to the reference: the Code git-history strip physically removes fix-bearing future commits, **including unreachable objects** (falling back to hiding `.git` only when the strip can't be verified); build-residue and cache scrubs; the answer-leak blocklist (mimoagent's list, extended with the Hub dataset hosts and package registries because this runs online where Xiaomi ran offline); a grader-isolation guard for in-container grading; the unprivileged `agent` user on Cyber and General. - Daytona backend support (`jobs/<kind>.daytona.yaml`): builds the Dockerfile and runs the agent unprivileged. - Music runs in a Debian *bookworm*-slim Python image with workdir `/root`, so it starts on HF Sandbox. ## Credits The environments, their prompts, tests, verifiers, images and graders are the work of the [Xiaomi MiMo](https://huggingface.co/XiaomiMiMo) team, released as [XiaomiMiMo/MiMo-V2.6-RL-oss](https://huggingface.co/datasets/XiaomiMiMo/MiMo-V2.6-RL-oss) together with their training code [XiaomiMiMo/verl](https://github.com/XiaomiMiMo/verl) and agent harness [XiaomiMiMo/mimoagent](https://github.com/XiaomiMiMo/mimoagent). mimoagent builds on [mini-swe-agent](https://github.com/SWE-agent/mini-swe-agent) (Kilian A. Lieret and Carlos E. Jimenez). The task format and runner are [Harbor](https://github.com/harbor-framework/harbor). This conversion is by Hugging Face and the FineEnvs team, and is not affiliated with or endorsed by Xiaomi. ## License Apache-2.0, as the source dataset (`LICENSE`). Files vendored from XiaomiMiMo/mimoagent (`tests/server_arvo.py` in Cyber tasks, and the ported setup commands) are MIT (`LICENSE-mimoagent.md`). `NOTICE` lists every source, its license, and the changes this conversion made. Vendored grader files are unmodified.