chanwit/gemma4-cub-agent-v7

---
license: apache-2.0
language:
- en
tags:
- gemma-4
- tool-calling
- reasoning
- thinking
- fine-tuning
- devops
- kubernetes
- confighub
- cub
- agent
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
dataset_info:
features:
- name: text
dtype: string
splits:
- name: train
num_bytes: 8822490
num_examples: 2032
download_size: 3023970
dataset_size: 8822490
---

# gemma4-cub-agent-v7

Training dataset for fine-tuning Gemma-4-31B-it as a DevOps agent with tool-calling and reasoning capabilities, specifically for managing Kubernetes clusters with [ConfigHub](https://confighub.com) (`cub` CLI).

## Why This Dataset

Fine-tuning Gemma-4 on reasoning-only data causes **catastrophic forgetting** of its native tool-calling. This dataset preserves both capabilities by mixing reasoning and tool-calling examples in Gemma-4's native token format.

## Composition

| Component | Examples | % |
|-----------|----------|---|
| Reasoning (coding, DevOps, kubectl, git) | 1362 | 72.6% |
| Tool-calling (cub, kubectl, argocd, flux) | 515 | 27.4% |
| **Total** | **1877** | |

## Stats

- **3184** tool calls across 515 examples
- **4456** thinking blocks across all examples
- **52** unique `cub` CLI commands verified against real ConfigHub server
- **All tool-calling examples** include `<|channel>thought` reasoning before every `<|tool_call>`
- Average **4,292** chars/example

## Tool-Calling Coverage

Covers 8 areas of ConfigHub management:
- **Unit lifecycle**: create, apply, refresh, diff, restore, approve, destroy
- **Functions**: get-replicas, set-image, yq-i, search-replace, vet-celexpr, etc.
- **Drift detection**: refresh, diff, livestate, reconciliation
- **GitOps**: ArgoCD + Flux via OCI bridge
- **Helm**: install, upgrade, template through ConfigHub
- **Workers**: create, install, status, logs
- **Multi-cluster**: push-upgrade, cross-space apply, changesets
- **Import**: adopt live resources, clean manifests, avoid SSA conflicts

Also covers kubectl troubleshooting (pod failures, OOMKill, CrashLoop, 502s, rollbacks), ArgoCD sync issues, Flux stuck kustomizations, and general coding tasks.

## Format

Native Gemma-4 tokens — ready for `SFTTrainer` with `train_on_responses_only`:

```
<bos><|turn>system
<|think|>System prompt<|tool>declaration:Bash{...}<tool|><turn|>
<|turn>user
User request<turn|>
<|turn>model
<|channel>thought
Step-by-step reasoning
<channel|><|tool_call>call:Bash{command:<|"|>kubectl get pods<|"|>}<tool_call|><turn|>
<|turn>tool
Command output<turn|>
<|turn>model
Final answer<turn|>
```

## Usage

```python
from datasets import load_dataset
from trl import SFTTrainer, SFTConfig

dataset = load_dataset("chanwit/gemma4-cub-agent-v7")

trainer = SFTTrainer(
model=model,
train_dataset=dataset["train"],
args=SFTConfig(
dataset_text_field="text",
max_seq_length=8192,
...
),
)

# Mask user/tool turns, train on model turns only
from trl import train_on_responses_only
trainer = train_on_responses_only(
trainer,
instruction_part="<|turn>user\n",
response_part="<|turn>model\n",
)
```

## Data Sources

- **Reasoning**: Filtered from Opus/Qwen reasoning datasets (coding, DevOps, kubectl, git focus)
- **Tool-calling**: Generated with Claude Opus, converted via hybrid Gemma-4 formatter
- **DevOps reasoning**: 30 kubectl/git/infrastructure entries generated with thinking blocks
- **All cub commands verified** by executing against real ConfigHub server + kind cluster

## License

Apache 2.0