这太重磅了。小米刚刚开源了超过 7000 个用于训练 MiMo 的强化学习任务环境，涵盖代码、网络安全、通用工具使用、可视化网页开发和音乐相关任务。以下是其意义超越又一个模型发布的原因：开放权重让你能运行模型，而开放环境让你能“教”出模型。强化学习环境难以构建，因为每个任务都需要一个可执行的世界、明确的目标、智能体可用的工具以及可靠的成功测试。小米发布了大量此前缺失的基础设施，包括提示词、智能体配置、奖励元数据以及配套的训练框架。开发者现在可以用它做什么？→ 在真实智能体任务上训练模型。 模型不再只是模仿静态答案，而是可以尝试任务、调用工具、观察结果并从结果中改进。→ 专门训练小型模型。 团队可以针对软件工程、漏洞复现、知识工作或网页开发专门训练一个开放检查点模型，而不是依赖一个大得多的通用模型。→ 研究完整轨迹。 每次运行都捕获了智能体的推理、工具调用、环境反馈和奖励。研究人员可以找出智能体在哪里失败，并比较训练算法、奖励策略、运行环境和模型。→ 构建完整的训练循环。 小米的技术栈涵盖了环境交互、轨迹收集、奖励评估和策略优化。开发者可以替换各个组件并衡量行为的改进。关键转变是从开源模型输出到开源学习经验。数据集教会模型答案“看起来”是什么样。而环境让模型能发现哪些行动能带来成功。在 HuggingFace 🤗 上查看：http://huggingface.co/datasets/XiaomiMiMo/MiMo-V2.6-RL-oss… 我写了一篇关于使用 GRPO 和 RULER 通过强化学习微调模型的完整解析。RULER 使用语言模型评判者来对轨迹进行排名，因此你无需手动编写奖励函数或收集标注答案。文章摘录如下。

XiaomiMiMo/MiMo-V2.6-RL-oss · Hugging Face 上的数据集

来源：https://huggingface.co/datasets/XiaomiMiMo/MiMo-V2.6-RL-oss [ { “content”: “[FEAT] Make `secrets` parameter optional so we can use this action just to get the token.\n**Is your feature request related to a problem? Please describe.**\r\nThe `secrets` parameter should not be mandatory. There is the option for `exportToken`, which is what I want to use to configure the Vaul... { “ground_truth”: “”, “style”: “rule” } { “dataset_type”: “opensource-code”, “index”: 1, “instance_id”: “format-code-task-001457”, “instance_json”: “{\“cwd\”: \“/testbed\”, \“dataset_type\”: \“opensource-code\”, \“docker_image\”: \“format-code-task-001457:latest\”, \“instance_id\”: \“format-code-task-001457\”, \“problem_statement\”: \“[FEAT] Make `se... [ { “content”: “I want `gz::sim::SdfEntityCreator` to support creating ECS entities from SDF projectors. Given an existing `SdfEntityCreator`, calling `Entity CreateEntities(const sdf::Projector *)` with a projector named `projector`, a raw pose of `(1, 2, 3, 0, 0, 0)`, near clip `2.0`, far clip `7.0`, horizontal... { “ground_truth”: “”, “style”: “rule” } { “dataset_type”: “opensource-code”, “index”: 2, “instance_id”: “format-code-task-001240”, “instance_json”: “{\“cwd\”: \“/workspace/repo\”, \“dataset_type\”: \“opensource-code\”, \“docker_image\”: \“format-code-task-001240:latest\”, \“instance_id\”: \“format-code-task-001240\”, \“problem_statement\”: \“I want `... [ { “content”: “## Role REST API silently drops `handle` on create / update\n\nI’m wiring up a small admin tool against the Corteza system API and I want to give each role a stable, machine-friendly `handle` next to its display `name` (similar to what users and the signup endpoint already accept). The `Role` type... { “ground_truth”: “”, “style”: “rule” } { “dataset_type”: “opensource-code”, “index”: 3, “instance_id”: “format-code-task-000723”, “instance_json”: “{\“cwd\”: \“/testbed\”, \“dataset_type\”: \“opensource-code\”, \“docker_image\”: \“format-code-task-000723:latest\”, \“instance_id\”: \“format-code-task-000723\”, \“problem_statement\”: \“## Role REST AP... [ { “content”: “# Problem Statement\n\n我现在在 Cortex 里上传 Alertmanager 配置时,receiver 的 HTTP 通知鉴权基本只能用 basic auth 或 bearer token,但我们这边的 webhook 接口要求走 OAuth2 client credentials。能不能让 `http_config` 这类通知配置直接支持 `oauth2`,比如用 client id、inline secret 和 token URL 去拿 token?另外这个配置是租户自己传的,最好不要允许它通过 `client_secret_file` 去读服务器上的本地文... { “ground_truth”: “”, “style”: “rule” } { “dataset_type”: “opensource-code”, “index”: 4, “instance_id”: “format-code-task-000718”, “instance_json”: “{\“cwd\”: \“/testbed\”, \“dataset_type\”: \“opensource-code\”, \“docker_image\”: \“format-code-task-000718:latest\”, \“instance_id\”: \“format-code-task-000718\”, \“problem_statement\”: \“# Problem State... [ { “content”: “## PI and LogEI acquisition functions crash when called\n\nI’m using RoBO for some Bayesian optimization experiments. The default `EI`\nacquisition function works fine, so I wanted to compare against `PI` and\n`LogEI` on the same problem.\n\nI set them up the same way I set up `EI` (same model, sa... { “ground_truth”: “”, “style”: “rule” } { “dataset_type”: “opensource-code”, “index”: 6, “instance_id”: “format-code-task-000484”, “instance_json”: “{\“cwd\”: \“/testbed\”, \“dataset_type\”: \“opensource-code\”, \“docker_image\”: \“format-code-task-000484:latest\”, \“instance_id\”: \“format-code-task-000484\”, \“problem_statement\”: \“## PI and LogEI... [ { “content”: “## BroydenSolver behaves inconsistently with other nonlinear solvers (label + recording)\n\nI’m running a model where I’m experimenting with different nonlinear solvers\n(BroydenSolver vs NewtonSolver vs NonlinearBlockGS) on the same group, and\nI have a case recorder attached so I can compare the... { “ground_truth”: “”, “style”: “rule” } { “dataset_type”: “opensource-code”, “index”: 7, “instance_id”: “format-code-task-000189”, “instance_json”: “{\“cwd\”: \“/testbed\”, \“dataset_type\”: \“opensource-code\”, \“docker_image\”: \“format-code-task-000189:latest\”, \“instance_id\”: \“format-code-task-000189\”, \“problem_statement\”: \“## BroydenSolve... [ { “content”: “## S3 bucket client keeps retrying after the request context is canceled\n\nWhen running Cortex against S3-backed object storage, I noticed that if an in-flight operation’s context gets canceled (either explicitly, or because an upstream deadline expires), the bucket client doesn’t actually stop r... { “ground_truth”: “”, “style”: “rule” } { “dataset_type”: “opensource-code”, “index”: 8, “instance_id”: “format-code-task-000720”, “instance_json”: “{\“cwd\”: \“/testbed\”, \“dataset_type\”: \“opensource-code\”, \“docker_image\”: \“format-code-task-000720:latest\”, \“instance_id\”: \“format-code-task-000720\”, \“problem_statement\”: \“## S3 bucket cl... [ { “content”: “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` s... { “ground_truth”: “”, “style”: “rule” } { “dataset_type”: “opensource-code”, “index”: 9, “instance_id”: “format-code-task-001661”, “instance_json”: “{\“cwd\”: \“/workspace/repo\”, \“dataset_type\”: \“opensource-code\”, \“docker_image\”: \“format-code-task-001661:latest\”, \“instance_id\”: \“format-code-task-001661\”, \“problem_statement\”: \“I want t... [ { “content”: “Dyson integration does not expose humidity or air quality (pm25) to homekit\n<!-- READ THIS FIRST:\r\n - If you need additional help with this template, please refer to https://www.home-assistant.io/help/reporting_issues/\r\n - Make sure you are running the latest version of Home Assistant befor... { “ground_truth”: “”, “style”: “rule” } { “dataset_type”: “opensource-code”, “index”: 10, “instance_id”: “format-code-task-001499”, “instance_json”: “{\“cwd\”: \“/testbed\”, \“dataset_type\”: \“opensource-code\”, \“docker_image\”: \“format-code-task-001499:latest\”, \“instance_id\”: \“format-code-task-001499\”, \“problem_statement\”: \“Dyson integrat... [ { “content”: “## Custom `dgs.graphql.graphiql.path` doesn’t actually work end-to-end\n\nI’m using the WebFlux variant of dgs-framework and I wanted to expose GraphiQL on a non-default path (we have a routing convention internally), so in my `application.yml` I set:\n\n```yaml\ndgs:\n graphql:\n graphiql:\n ... { “ground_truth”: “”, “style”: “rule” } { “dataset_type”: “opensource-code”, “index”: 11, “instance_id”: “format-code-task-000173”, “instance_json”: “{\“cwd\”: \“/testbed\”, \“dataset_type\”: \“opensource-code\”, \“docker_image\”: \“format-code-task-000173:latest\”, \“instance_id\”: \“format-code-task-000173\”, \“problem_statement\”: \“## Custom `dgs... [ { “content”: “pvsystem.Array.get_irradiance raises an error if solar_zenith is passed as a float\n**Bug Description**\r\n`pvsystem.get_irradiance` documentation says solar_zenith can be passed as a float or a Series. However, if `dni_extra` is not defined and `solar_zenith` is passed as float, then `solar_zenit... { “ground_truth”: “”, “style”: “rule” } { “dataset_type”: “opensource-code”, “index”: 12, “instance_id”: “format-code-task-002366”, “instance_json”: “{\“cwd\”: \“/testbed\”, \“dataset_type\”: \“opensource-code\”, \“docker_image\”: \“format-code-task-002366:latest\”, \“instance_id\”: \“format-code-task-002366\”, \“problem_statement\”: \“pvsystem.Array... [ { “content”: “ufuzz failure\n```js\r\n// original code\r\n// (beautified)\r\nvar _calls_ = 10, a = 100, b = 10, c = 0;\r\n\r\nfunction f0(arguments_1, a_1) {\r\n {\r\n var brake1 = 5;\r\n L12709: while ((c = c + 1) + a++ && --brake1 > 0) {\r\n try {\r\n } catch (undefined_... { “ground_truth”: “”, “style”: “rule” } { “dataset_type”: “opensource-code”, “index”: 13, “instance_id”: “format-code-task-001933”, “instance_json”: “{\“cwd\”: \“/testbed\”, \“dataset_type\”: \“opensource-code\”, \“docker_image\”: \“format-code-task-001933:latest\”, \“instance_id\”: \“format-code-task-001933\”, \“problem_statement\”: \“ufuzz failure\... [ { “content”: “I want the ReNative CLI to provide hook-management commands for project build hooks: `rnv hooks run`, `rnv hooks list`, and `rnv hooks pipes`.\n\n`rnv hooks run -x ` and `rnv hooks run --exe-method ` should configure the project when a project config exists, load the project bu... { “ground_truth”: “”, “style”: “rule” } { “dataset_type”: “opensource-code”, “index”: 14, “instance_id”: “format-code-task-001190”, “instance_json”: “{\“cwd\”: \“/workspace/repo\”, \“dataset_type\”: \“opensource-code\”, \“docker_image\”: \“format-code-task-001190:latest\”, \“instance_id\”: \“format-code-task-001190\”, \“problem_statement\”: \“I want ...