gemma-4-e2b-kubectl-mcp-server-mcp-sft-v2-it

MCP tool-calling 的 完整微調模型（full fine-tune），由 Agent Tools Fine-Tuning Platform 以 「Teacher 反向生成資料 → 品質過濾 → SFT」流程訓練而成。

• Base model：google/gemma-4-E2B-it
• 方法：Full SFT（epochs=3.0, learning_rate=1e-05）
• 工具（來自 MCP server）：backup_create_tool, audit_rbac_permissions, certs_issuers_list_tool, istio_destinationrules_list_tool, helm_repo_update, helm_show_readme, helm_dependency_update, keda_scaledobjects_list_tool, keda_detect_tool, certs_issuer_get_tool, kind_ingress_setup_tool, policy_list_tool, kubevirt_vm_restart_tool, compare_namespaces, analyze_pod_security, kubevirt_vmis_list_tool, taint_node, vind_pause_tool, vind_create_cluster_tool, helm_get_all, rollout_get_tool, node_top_tool, backup_detect_tool, kind_config_show_tool, certs_challenges_list_tool, kind_node_restart_tool, get_logs, helm_rollback, install_helm_chart, backup_get_tool, policy_detect_tool, gitops_detect_engine_tool, vind_resume_tool, helm_version, kind_build_node_image_tool, kind_registry_status_tool, get_ingress, kind_config_validate_tool, kind_export_logs_tool, get_daemonsets, vind_logs_tool, istio_detect_tool, vind_platform_start_tool, diagnose_pod_crash, keda_scaledobject_get_tool, get_pdb, cilium_detect_tool, get_crds, helm_package, helm_show_chart, helm_create, keda_triggerauth_get_tool, kubevirt_datavolumes_list_tool, kind_node_logs_tool, capi_clusters_list_tool, kubectl_apply, get_api_resources, helm_search_repo, vind_upgrade_tool, get_idle_resources, istio_authorizationpolicies_list_tool, vind_detect_tool, get_jobs, kubevirt_vm_start_tool, helm_list, kind_node_inspect_tool, kubevirt_instancetypes_list_tool, get_service_accounts, capi_machinedeployments_list_tool, istio_sidecar_status_tool, kind_set_kubeconfig_tool, helm_get_hooks, get_pvcs, kind_get_kubeconfig_tool, kind_available_images_tool, kubevirt_vm_stop_tool, vind_describe_tool, certs_status_explain_tool, get_resource_quotas, kubevirt_vm_unpause_tool, multi_cluster_query, kubevirt_datasources_list_tool, run_pod, get_evicted_pods, get_admission_webhooks, get_overprovisioned_resources, capi_machine_get_tool, istio_virtualservice_get_tool, kind_network_inspect_tool, optimize_resource_requests, kubevirt_vms_list_tool, capi_machinedeployment_scale_tool, get_endpoints, kubectl_explain, kubectl_create, cilium_status_tool, get_storage_classes, get_cluster_info, helm_search_hub, get_replicasets, vind_list_clusters_tool, flagger_canary_get_tool, gitops_source_get_tool, get_rbac_roles, helm_history, gitops_app_sync_tool, kubeconfig_view, gitops_app_get_tool, rollout_promote_tool, helm_show_all, multi_cluster_health, node_logs_tool, gitops_app_status_tool, cilium_nodes_list_tool, diagnose_network_connectivity, get_configmaps, node_stats_summary_tool, kind_node_labels_tool, istio_virtualservices_list_tool, get_api_versions, backup_schedules_list_tool, backup_locations_list_tool, kubevirt_vm_pause_tool, kind_list_clusters_tool, helm_pull, detect_pending_pods, get_limit_ranges, capi_machinehealthchecks_list_tool, helm_dependency_build, get_pod_security_info, istio_analyze_tool, get_pods, keda_hpa_list_tool, get_events, set_namespace_for_context, switch_context, kind_cluster_status_tool, capi_machinesets_list_tool, restore_list_tool, istio_proxy_status_tool, list_contexts_tool, certs_requests_list_tool, cilium_policy_get_tool, get_namespaces, kubectl_cp, restart_deployment, capi_clusterclasses_list_tool, helm_status, kind_get_nodes_tool, capi_detect_tool, vind_connect_tool, kubectl_describe, wait_for_condition, scale_deployment, delete_resource, get_resource_usage, helm_repo_list, multi_cluster_pod_count, disable_kubeconfig_watching, backup_delete_tool, keda_scaledjobs_list_tool, kind_load_image_tool, helm_test, kind_images_list_tool, check_secrets_security, capi_cluster_get_tool, capi_machines_list_tool, check_crd_exists, enable_kubeconfig_watching, helm_template_apply, vind_disconnect_tool, get_server_config_status, get_resource_recommendations, node_management, exec_in_pod, helm_show_values, kubectl_patch, capi_cluster_kubeconfig_tool, policy_audit_tool, kind_create_cluster_tool, certs_renew_tool, kubevirt_detect_tool, istio_gateways_list_tool, helm_show_crds, port_forward, get_previous_logs, kubevirt_vm_migrate_tool, annotate_resource, cleanup_pods, upgrade_helm_chart, certs_get_tool, helm_template, check_pod_health, kind_node_exec_tool, kind_delete_all_clusters_tool, trace_service_chain, backup_resource, helm_repo_remove, uninstall_helm_chart, get_hpa, istio_peerauthentications_list_tool, get_statefulsets, get_cluster_roles, kind_detect_tool, vind_delete_cluster_tool, cilium_identities_list_tool, kind_port_mappings_tool, helm_lint, kubectl_rollout, label_resource, rollout_restart_tool, kind_cluster_info_tool, rollout_abort_tool, kind_load_image_archive_tool, helm_get_values, keda_triggerauths_list_tool, get_node_metrics, restore_get_tool, policy_get_tool, get_secrets, certs_detect_tool, get_namespace_cost_allocation, get_pod_metrics, get_context_details, get_persistent_volumes, get_resource_quotas_usage, cilium_policies_list_tool, certs_list_tool, kubevirt_vm_get_tool, create_deployment, get_nodes_summary, hubble_flows_query_tool, restore_create_tool, analysis_runs_list_tool, get_resource_trends, get_priority_classes, vind_get_kubeconfig_tool, helm_env, kind_delete_cluster_tool, rollouts_detect_tool, vind_status_tool, kind_provider_info_tool, helm_get_manifest, health_check, rollouts_list_tool, backup_list_tool, kind_registry_create_tool, helm_dependency_list, get_services, helm_repo_add, policy_explain_denial_tool, kubectl_generic, get_cost_analysis, kind_version_tool, get_cluster_version, backup_schedule_create_tool, kind_config_generate_tool, gitops_apps_list_tool, policy_violations_list_tool, gitops_sources_list_tool, set_server_stateless_mode, kind_registry_connect_tool, get_nodes, check_dns_resolution, get_pod_events, cilium_endpoints_list_tool, rollout_retry_tool, get_current_context, flagger_canaries_list_tool, get_pod_conditions, list_crds, analyze_network_policies, rollout_status_tool, helm_get_notes, get_deployments

評估（held-out test）

指標	Base（訓練前）	Fine-tuned（訓練後）
有工具呼叫率	100.0%	99.8%
函式名正確率	86.2%	88.8%
名稱+參數全對率	74.7%	84.1%
負樣本正確率	98.2%	100.0%

使用方式

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Simon-Liu/gemma-4-e2b-kubectl-mcp-server-mcp-sft-v2-it", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Simon-Liu/gemma-4-e2b-kubectl-mcp-server-mcp-sft-v2-it")

模型被訓練成：當使用者的請求需要工具時，輸出單一 <tool_call>{"name": "<tool>", "arguments": {...}}</tool_call>；不需要工具時，直接以純文字回答。