README.md

---
library_name: mlx
license: apache-2.0
license_link: https://ai.google.dev/gemma/docs/gemma_4_license
base_model: google/gemma-4-E2B-it
datasets:
  - htunn/aiops-gemma
language:
  - en
tags:
  - mlx
  - gemma4
  - lora
  - aiops
  - kubernetes
  - nutanix
  - vmware
  - active-directory
  - pki
  - infrastructure-automation
  - conversational
pipeline_tag: text-generation
---

# Gemma 4 E2B — AIOps Orchestrator (Safetensors / HF format)

Fine-tuned from **[google/gemma-4-E2B-it](https://huggingface.co/google/gemma-4-E2B-it)** using Apple MLX LoRA on an M3 Pro (18 GB). The model is specialised as an **autonomous AIOps orchestrator agent** that maps infrastructure incident telemetry to structured, execution-ready JSON control schemas.

> Looking for the quantised GGUF variant?  
> → **[htunn/gemma-4-e2b-aiops-gguf](https://huggingface.co/htunn/gemma-4-e2b-aiops-gguf)**

---

## Model Details

| Field | Value |
|---|---|
| **Base model** | `google/gemma-4-E2B-it` |
| **Architecture** | `Gemma4ForConditionalGeneration` |
| **Precision** | BF16 |
| **Parameters** | ~5B |
| **Fine-tune method** | LoRA (MLX) |
| **LoRA rank / scale** | 8 / 20.0 |
| **Layers tuned** | 16 |
| **Training iterations** | 600 |
| **Batch size** | 1 |
| **Learning rate** | 1e-4 |
| **Max sequence length** | 2048 |
| **Training hardware** | Apple M3 Pro, 18 GB Unified Memory |
| **Framework** | `mlx-lm` |
| **Dataset** | [htunn/aiops-gemma](https://huggingface.co/datasets/htunn/aiops-gemma) |
| **License** | Apache 2.0 |

---

## Intended Use

The model receives an `[AIOps-Agent]` incident description spanning one or more infrastructure domains and returns a **single JSON object** containing:

- `action` — the remediation action to take
- `target_*` — specific resource identifiers
- `steps` — ordered list of discrete remediation steps
- `api_calls` — REST API calls to execute
- `priority` / `risk` — severity classification

### Covered Domains

- **Kubernetes** — pod failures, etcd quorum loss, network partitions, cert-manager, autoscaling, storage
- **Nutanix** — CVM network partitions, multi-cluster Prism Central, NearSync/async replication, Calm blueprints, Karbon, Era, stretched clusters, Flow microsegmentation
- **VMware** — ESXi, vSAN, NSX-T, DRS, HA failover, vMotion, vCenter SSO
- **Active Directory** — replication, FSMO, SYSVOL, cross-site failures, forest trusts
- **ADFS / PKI** — certificate chain validation, OCSP, CRL, intermediate CA renewal, MFA providers
- **Windows Server** — domain controllers, IIS, WSUS, Hyper-V, DFS, file services

---

## Usage

### Transformers

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "htunn/gemma-4-e2b-aiops-hf"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

prompt = "[AIOps-Agent] Kubernetes etcd leader election failed due to network partition isolating etcd-01. API server returning 503. etcd cluster lost quorum."

messages = [{"role": "user", "content": prompt}]
inputs = tokenizer.apply_chat_template(
    messages,
    return_tensors="pt",
    add_generation_prompt=True,
).to(model.device)

outputs = model.generate(inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))
```

### MLX (Apple Silicon)

```bash
pip install mlx-lm
mlx_lm.generate \
  --model htunn/gemma-4-e2b-aiops-hf \
  --prompt "[AIOps-Agent] Nutanix stretched cluster witness unreachable. RF1 degraded mode active." \
  --max-tokens 256
```

---

## Example Output

**Input:**
```
[AIOps-Agent] Kubernetes ingress controller presenting incomplete TLS certificate chain.
Intermediate CA missing from secret bundle. External clients receiving SSL handshake errors.
```

**Output:**
```json
{
  "action": "fix_incomplete_cert_chain",
  "target_service": "ingress-nginx",
  "missing_element": "intermediate_ca_certificate",
  "steps": [
    "retrieve_full_certificate_chain_from_pki",
    "bundle_leaf_cert_with_intermediate_ca_cert",
    "update_kubernetes_tls_secret_with_complete_chain",
    "rolling_restart_ingress_controller_pods",
    "validate_chain_with_openssl_s_client"
  ],
  "commands": [
    "openssl s_client -connect ingress.example.com:443 -showcerts",
    "openssl verify -CAfile chain.pem cert.pem"
  ],
  "api_calls": [
    "PATCH /api/v1/namespaces/ingress-nginx/secrets/tls-wildcard-cert"
  ],
  "priority": "high"
}
```

---

## Training Data

Trained on **[htunn/aiops-gemma](https://huggingface.co/datasets/htunn/aiops-gemma)** — 90 hand-crafted incident scenarios (73 train / 17 validation) in Gemma chat format, covering:

- Network partition events across K8s, Nutanix, AD, and NSX-T
- Certificate chain validation failures (incomplete chains, expired intermediates, OCSP, path-length constraints)
- Multi-cluster Nutanix operations (Prism Central, NearSync, Calm, Era, Karbon, stretched clusters)
- Cross-domain cascading failures involving 3–4 infrastructure layers simultaneously

---

## Limitations

- Output format is tailored to the training schema; prompts not prefixed with `[AIOps-Agent]` may produce inconsistent results.
- The model does not execute actions — it produces decision schemas for an orchestration layer to consume.
- Coverage is limited to the incident types represented in the training set.

---

## Related Repos

| Repo | Description |
|---|---|
| [htunn/gemma-4-e2b-aiops-gguf](https://huggingface.co/htunn/gemma-4-e2b-aiops-gguf) | Q4_K_M GGUF — run with Ollama or llama.cpp |
| [htunn/aiops-gemma](https://huggingface.co/datasets/htunn/aiops-gemma) | Training dataset (JSONL) |
| [GitHub: htunn/aiops-gemma4](https://github.com/htunn/aiops-gemma4) | Full fine-tuning pipeline source |