Simon-Liu/kubectl-mcp-server-tool-call-reasoning-6k

kubectl-mcp-server-tool-call-reasoning-6k

MCP tool-calling SFT 資料集，由 Agent Tools Fine-Tuning Platform 以「反向生成 + teacher solver 驗證」流程產生。

• —語言：繁體中文
• —工具（來自 MCP server）：install_helm_chart, upgrade_helm_chart, uninstall_helm_chart, helm_list, helm_status, helm_history, helm_get_values, helm_get_manifest, helm_get_notes, helm_get_hooks, helm_get_all, helm_show_chart, helm_show_values, helm_show_readme, helm_show_crds, helm_show_all, helm_search_repo, helm_search_hub, helm_repo_list, helm_repo_add, helm_repo_remove, helm_repo_update, helm_rollback, helm_test, helm_lint, helm_package, helm_dependency_update, helm_dependency_list, helm_dependency_build, helm_pull, helm_create, helm_version, helm_env, helm_template, helm_template_apply, get_pods, get_logs, get_pod_events, check_pod_health, exec_in_pod, cleanup_pods, run_pod, get_pod_conditions, get_previous_logs, diagnose_pod_crash, detect_pending_pods, get_evicted_pods, get_namespaces, get_services, get_nodes, get_configmaps, get_secrets, get_events, get_resource_usage, get_service_accounts, get_crds, get_resource_quotas, get_limit_ranges, get_priority_classes, list_contexts_tool, get_current_context, get_context_details, kubeconfig_view, switch_context, set_namespace_for_context, get_cluster_info, get_cluster_version, health_check, kubectl_explain, get_api_resources, get_api_versions, get_admission_webhooks, check_crd_exists, list_crds, get_nodes_summary, get_server_config_status, enable_kubeconfig_watching, disable_kubeconfig_watching, set_server_stateless_mode, node_logs_tool, node_stats_summary_tool, node_top_tool, multi_cluster_query, multi_cluster_health, multi_cluster_pod_count, get_deployments, create_deployment, scale_deployment, restart_deployment, get_replicasets, get_statefulsets, get_daemonsets, get_jobs, get_hpa, get_pdb, get_rbac_roles, get_cluster_roles, analyze_pod_security, analyze_network_policies, audit_rbac_permissions, check_secrets_security, get_pod_security_info, get_ingress, get_endpoints, diagnose_network_connectivity, check_dns_resolution, trace_service_chain, port_forward, get_persistent_volumes, get_pvcs, get_storage_classes, kubectl_apply, kubectl_describe, kubectl_generic, kubectl_patch, kubectl_rollout, kubectl_create, delete_resource, kubectl_cp, backup_resource, label_resource, annotate_resource, taint_node, wait_for_condition, node_management, compare_namespaces, get_pod_metrics, get_node_metrics, get_resource_recommendations, get_idle_resources, get_resource_quotas_usage, get_cost_analysis, get_overprovisioned_resources, get_resource_trends, get_namespace_cost_allocation, optimize_resource_requests, gitops_apps_list_tool, gitops_app_get_tool, gitops_app_sync_tool, gitops_app_status_tool, gitops_sources_list_tool, gitops_source_get_tool, gitops_detect_engine_tool, certs_list_tool, certs_get_tool, certs_issuers_list_tool, certs_issuer_get_tool, certs_renew_tool, certs_status_explain_tool, certs_challenges_list_tool, certs_requests_list_tool, certs_detect_tool, policy_list_tool, policy_get_tool, policy_violations_list_tool, policy_explain_denial_tool, policy_audit_tool, policy_detect_tool, backup_list_tool, backup_get_tool, backup_create_tool, backup_delete_tool, restore_list_tool, restore_create_tool, restore_get_tool, backup_locations_list_tool, backup_schedules_list_tool, backup_schedule_create_tool, backup_detect_tool, keda_scaledobjects_list_tool, keda_scaledobject_get_tool, keda_scaledjobs_list_tool, keda_triggerauths_list_tool, keda_triggerauth_get_tool, keda_hpa_list_tool, keda_detect_tool, cilium_policies_list_tool, cilium_policy_get_tool, cilium_endpoints_list_tool, cilium_identities_list_tool, cilium_nodes_list_tool, cilium_status_tool, hubble_flows_query_tool, cilium_detect_tool, rollouts_list_tool, rollout_get_tool, rollout_status_tool, rollout_promote_tool, rollout_abort_tool, rollout_retry_tool, rollout_restart_tool, analysis_runs_list_tool, flagger_canaries_list_tool, flagger_canary_get_tool, rollouts_detect_tool, capi_clusters_list_tool, capi_cluster_get_tool, capi_machines_list_tool, capi_machine_get_tool, capi_machinedeployments_list_tool, capi_machinedeployment_scale_tool, capi_machinesets_list_tool, capi_machinehealthchecks_list_tool, capi_clusterclasses_list_tool, capi_cluster_kubeconfig_tool, capi_detect_tool, kubevirt_vms_list_tool, kubevirt_vm_get_tool, kubevirt_vmis_list_tool, kubevirt_vm_start_tool, kubevirt_vm_stop_tool, kubevirt_vm_restart_tool, kubevirt_vm_pause_tool, kubevirt_vm_unpause_tool, kubevirt_vm_migrate_tool, kubevirt_datasources_list_tool, kubevirt_instancetypes_list_tool, kubevirt_datavolumes_list_tool, kubevirt_detect_tool, istio_virtualservices_list_tool, istio_virtualservice_get_tool, istio_destinationrules_list_tool, istio_gateways_list_tool, istio_peerauthentications_list_tool, istio_authorizationpolicies_list_tool, istio_proxy_status_tool, istio_analyze_tool, istio_sidecar_status_tool, istio_detect_tool, vind_detect_tool, vind_list_clusters_tool, vind_status_tool, vind_get_kubeconfig_tool, vind_logs_tool, vind_create_cluster_tool, vind_delete_cluster_tool, vind_pause_tool, vind_resume_tool, vind_connect_tool, vind_disconnect_tool, vind_upgrade_tool, vind_describe_tool, vind_platform_start_tool, kind_detect_tool, kind_version_tool, kind_list_clusters_tool, kind_get_nodes_tool, kind_get_kubeconfig_tool, kind_export_logs_tool, kind_cluster_info_tool, kind_node_labels_tool, kind_create_cluster_tool, kind_delete_cluster_tool, kind_delete_all_clusters_tool, kind_load_image_tool, kind_load_image_archive_tool, kind_build_node_image_tool, kind_set_kubeconfig_tool, kind_config_validate_tool, kind_config_generate_tool, kind_config_show_tool, kind_available_images_tool, kind_registry_create_tool, kind_registry_connect_tool, kind_registry_status_tool, kind_node_exec_tool, kind_node_logs_tool, kind_node_inspect_tool, kind_node_restart_tool, kind_network_inspect_tool, kind_port_mappings_tool, kind_ingress_setup_tool, kind_cluster_status_tool, kind_images_list_tool, kind_provider_info_tool
• —Train / Test：5005 / 556
• —負樣本比例：0.1
• —Teacher 資料品質：函式名正確率 0.9908、名稱+參數全對率 0.9049

欄位

生成方式

先從 MCP 工具 schema 抽出「目標工具 + 合法參數」作為 ground truth，再由 teacher model 生成對應的自然語言使用者問題；另由 teacher 以 solver 身分重新解題做品質過濾（只保留還原正確者）。因此 expected_tool_calls 可作為可精確比對的標準答案，用來計算有工具呼叫率 / 函式名正確率 / 名稱+參數全對率。

載入

from datasets import load_dataset
ds = load_dataset("Simon-Liu/kubectl-mcp-server-tool-call-reasoning-6k")
print(ds["train"][0])