FineEnvs/MiMo-V2.6-RL-harbor-webdev

MiMo-V2.6-RL Webdev (Harbor)

Build a website from a design brief. 2,093 Harbor tasks from the Webdev domain of Xiaomi's 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 builds the site described in the brief and delivers it to dist/. The grader renders the page in a headless browser and a vision model scores the full-page screenshot on layout, typography, colour, whitespace, content, brief fulfilment and assets.

Tasks	2,093
Graded by	vision judge (Llama-4-Maverick)
Reference step limit	64
Source	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, cyber, general, terminal, webdev, music.

Layout

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/webdev.yaml      # the reference agent settings
agents/mimo_opencode.py

Example: task.toml · instruction · verifier.

Run

hf download FineEnvs/MiMo-V2.6-RL-harbor-webdev --repo-type dataset --local-dir mimo-webdev
cd mimo-webdev
PYTHONPATH=agents HF_TOKEN=hf_... harbor run -y -c jobs/webdev.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/webdev.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:

openenv harbor serve --dataset FineEnvs/MiMo-V2.6-RL-harbor-webdev --llm-url http://127.0.0.1:8000/v1 --model <your-model>

How it was converted

With the mimo_harbor adapter. 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, 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.
• Grading can't be gamed. Xiaomi grades in a separate container; we grade in the same one, but first throw away anything the agent might have left to fake a score (a planted result file, leftover processes, planted Python start-up files), so only the real grader sets the reward.
• The vision judge is fixed to meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8 (Xiaomi used whatever its training run had). Override it with MIMO_VISION_JUDGE. For the strongest grading isolation, run on Daytona (it grades in its own container); on HF Sandbox grading shares the agent's container.

That's the whole list. The detailed, technical version is in the adapter README.

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, 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)	0.749 ± 0.049	0.650, 0.803, 0.793
  Harbor	0.738 ± 0.054	0.799, 0.631, 0.785

  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 team, released as XiaomiMiMo/MiMo-V2.6-RL-oss together with their training code XiaomiMiMo/verl and agent harness XiaomiMiMo/mimoagent. mimoagent builds on mini-swe-agent (Kilian A. Lieret and Carlos E. Jimenez). The task format and runner are 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.