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

gemma-4-e2b-kubectl-mcp-server-mcp-sft-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=2e-06）
• 工具（來自 MCP server）：get_deployments, scale_deployment, install_helm_chart, upgrade_helm_chart, uninstall_helm_chart, get_pods, detect_pending_pods, get_nodes, get_nodes_summary, diagnose_pod_crash, get_logs, helm_get_values, helm_list, trace_service_chain, helm_search_repo, get_pod_metrics, list_certs, kubectl_rollout, get_previous_logs, exec_in_pod, list_crds, discover_crds, create_deployment, get_idle_resources, helm_repo_list, helm_repo_update, get_policy_violations, get_jobs, get_pod_events, helm_rollback, get_ingress, kubectl_describe, helm_history, get_resource_usage, get_resource_recommendations, helm_status, node_stats_summary, health_check, kubectl_cp, diagnose_network_connectivity, get_resource_quotas, switch_context, get_cluster_info, detect_certs, list_cert_issuers, get_node_metrics, get_pvcs, optimize_resource_requests, check_dns_resolution, audit_rbac_permissions, get_services, get_endpoints, search_crds, list_backups, create_restore, get_cert, multi_cluster_pod_count, explain_policy_denial, get_storage_classes, restart_deployment, get_evicted_pods, cleanup_pods, gitops_apps_list, check_pod_health, list_custom_resources, get_resource_quotas_usage, analyze_network_policies, get_configmaps, compare_namespaces, get_pod_conditions, helm_version, get_namespaces, get_current_context, list_contexts, get_server_config_status, get_service_accounts, kubectl_explain, set_namespace_for_context, get_rollout_status, node_top, run_pod, explain_cert_status, get_overprovisioned_resources, taint_node, helm_show_values, check_crd_exists, helm_template, get_cluster_version, get_limit_ranges, delete_resource, get_rbac_roles, gitops_app_status, cilium_list_policies, node_management, helm_repo_add, port_forward, get_persistent_volumes, get_rollouts_list, get_api_resources, detect_rollouts, detect_backup, list_backup_schedules, cilium_list_identities, cilium_get_status, detect_policy_engines, get_policy_list, list_cert_challenges, get_events, analyze_pod_security, get_secrets, get_hubble_flows, promote_rollout, describe_crd, get_custom_resource, check_secrets_security, helm_search_hub, istio_sidecar_status, istio_analyze, get_hpa, multi_cluster_health, get_resource_trends, get_daemonsets, get_context_details, get_cost_analysis, helm_show_readme, kubevirt_vmis_list, gitops_app_sync, get_crds, list_cert_requests, get_pdb, multi_cluster_query, helm_get_all, renew_cert, vind_create_cluster, keda_detect, keda_scaledobjects_list, kubevirt_vms_list, get_rollout, gitops_detect_engine, wait_for_condition, kind_cluster_status, backup_resource, helm_get_manifest, helm_repo_remove, node_logs, helm_show_all, kind_cluster_info, gitops_app_get, kind_registry_status, kubectl_create, vind_get_kubeconfig, kubectl_patch

評估（held-out test）

使用方式

from transformers import AutoModelForCausalLM, AutoTokenizer

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