chanwit/gemma4-cub-agent-v11
收藏资源简介:
--- license: apache-2.0 language: - en tags: - gemma-4 - tool-calling - agent - confighub - kubernetes - devops - reasoning size_categories: - 1K<n<10K --- # gemma4-cub-agent-v11 Fine-tuning dataset for a Gemma-4 based ConfigHub/Kubernetes DevOps agent. ## What's new in v11 Builds on v10 (chanwit/gemma4-cub-agent-v10) with: - **Schema-verified command correctness** — every `cub` command in the dataset validated against the real cobra command tree from `github.com/confighub/sdk`. A stripped-down stub-cub fork was built to extract the full flag/arg/subcommand schema (266 commands) which was then used to auto-validate every training example. - **Known bug classes fixed:** - 21 × `cub k8s source ... --space X` → `--namespace X` (k8s source queries live K8s via kubeconfig, not a ConfigHub space) - 6 × `cub unit tag` positional ordering (correct form: `cub unit tag <tag> --unit <unit>`) - 6 dropped entries with unrecoverable positional-arg mismatches - **Expanded coverage** (new in v11): - 140 value-variation drills (set-replicas N, set-container-image, set-env-var, set-annotation, changeset --description, tag create) with varying values - 60 full 4-step workflow examples (`cub k8s source` → `--help` → `cub function do` → `cub unit apply`) - 80 reasoning/action-matched examples (thinking block explicitly names the command it generates) - 170 flag-pattern drills across underrepresented subcommands - 25 vocabulary mappings (umbrella terms → concrete kubectl kinds) - 55 counter-examples disambiguating `cub unit get` metadata vs `cub function do get-*` workload state - **Regression coverage:** 25/25 known bug classes from v5-v10 verified present in training data via `scripts/eval/sanity_check_dataset.py`. ## Statistics - **Total entries**: 7,114 - **Tool-calling entries**: 1,774 (25%) - **Reasoning entries**: 5,340 (75%) - **Total cub commands**: 4,837 - **Invalid commands**: 0 - **Format**: Gemma-4 text-only (`<|turn>system / user / model / tool<turn|>`) ## Intended training config - Base: `unsloth/gemma-3-27b-it-bnb-4bit` or equivalent Gemma-4 31B - LoRA: r=16, alpha=16 - Epochs: 2 - Learning rate: 2e-5 - train_on_responses_only ## License Apache 2.0




