xayrullonematov/hamma-data

HAMMA DevOps Failure States Dataset ⬛⬜

This dataset was created to fine-tune the core intelligence engine for HAMMA — a local-first, zero-cloud SSH client built for the Gemma 4 Good Hackathon. It contains 3,701 curated problem-solution pairs covering the full surface area of Linux server administration: permissions, systemd, networking, Docker, Kubernetes, databases, storage, security, and CI/CD.

---

🎯 The "Zero-Fluff" Philosophy

Standard instruction-tuned LLMs respond to server errors with conversational padding:

"I'd be happy to help you with that Nginx error! It looks like you may have a port collision on 0.0.0.0:80. Let me walk you through some troubleshooting steps..."

In a live production incident, engineers need the exact command — not a preamble. HAMMA is designed to work inside a terminal SSH client where every token of padding is noise.

This dataset is structured to train that behavior out. Every assistant response follows a strict format:

1. Root cause — one to three sentences of plain-English diagnosis
2. Exact fix — copy-pasteable commands in fenced code blocks
3. Verification — how to confirm the fix worked

No filler. No "great question!". No "I hope this helps."

---

📊 Dataset Statistics

Metric	Value
Total examples	3,701
Unique topics	39
Sub-angles covered	273
Avg user message length	~365 chars
Avg assistant response length	~1,467 chars
Responses with code blocks	100%
Duplicate prompts	0
Format	Chat (messages) + Prompt/Completion

Topic Distribution (top 10)

Topic	Examples
systemd unit failures and service debugging	140
Cron and scheduled job silent failures	140
SSH Permission denied (publickey)	140
SFTP chroot jail and access issues	140
Docker container exit codes and crash debugging	140
iptables and nftables firewall debugging	139
Docker bridge network collisions	139
Docker image build and layer cache failures	139
SELinux and AppArmor denials	137
MTU mismatches and packet fragmentation	137

View all 39 topics

• File permission and ownership errors
• systemd unit failures and service debugging
• Cron and scheduled job silent failures
• SELinux and AppArmor denials
• Kernel panics and dmesg analysis
• DNS resolution failures
• iptables and nftables firewall debugging
• TCP connection problems (refused, reset, timeout)
• TLS/SSL certificate problems
• MTU mismatches and packet fragmentation
• SSH Permission denied (publickey)
• SFTP chroot jail and access issues
• Docker bridge network collisions
• Docker container exit codes and crash debugging
• Docker image build and layer cache failures
• Docker volume mount and permission issues
• Pod CrashLoopBackOff debugging
• ImagePullBackOff and registry auth
• PersistentVolumeClaim stuck in Pending
• Kubernetes Ingress and Service debugging
• Node NotReady and kubelet issues
• Nginx upstream and 502/504 errors
• Apache .htaccess and mod_rewrite failures
• HAProxy backend health and routing
• PostgreSQL connection limits and pgbouncer
• MySQL replication broken
• Redis OOM and eviction issues
• Disk full with no obvious culprit
• NFS mount hangs and stale file handles
• Linux OOM killer traces and prevention
• High load average with low CPU usage (iowait)
• SSH brute force and fail2ban
• Let's Encrypt / certbot renewal failures
• AWS EC2 instance status checks failing
• S3 access denied debugging
• Git operational issues (conflicts, detached HEAD, large files)
• CI/CD pipeline runner failures
• Log analysis with journalctl, grep, awk
• Time sync issues (NTP/chrony)

---

📁 Dataset Structure

The dataset ships in two files covering the same 3,701 examples in different schemas:

hamma_devops_messages.jsonl — Chat format (recommended)

Use this for Gemma 4 / Unsloth / SFTTrainer fine-tuning. Each line is:

{
  "messages": [
    {
      "role": "system",
      "content": "You are a senior DevOps and Linux server engineer. Diagnose root causes precisely from logs, errors, and configs. Provide exact, runnable terminal commands and configuration. Warn before destructive operations. Prefer modern tooling."
    },
    {
      "role": "user",
      "content": "After upgrading sudo package on Ubuntu 22.04, users can't run ping anymore:\n\n$ ping 8.8.8.8\nping: socktype: SOCK_RAW: Operation not permitted\nping: socket: Operation not permitted\nping: => missing cap_net_raw+p capability or setuid?\n\nls -l /usr/bin/ping shows:\n-rwxr-xr-x 1 root root 72776 Feb 16 2023 /usr/bin/ping\n\nWhat happened?"
    },
    {
      "role": "assistant",
      "content": "The setuid bit was stripped from `/usr/bin/ping` during the package upgrade. Ping requires either CAP_NET_RAW capability or the setuid bit to create raw sockets. The expected permissions should be `-rwsr-xr-x` (note the `s` in owner execute position).\n\nRestore the setuid bit:\n\n```bash\nsudo chmod u+s /usr/bin/ping\n```\n\nVerify the fix:\n\n```bash\nls -l /usr/bin/ping\nping -c 1 8.8.8.8\n```\n\nYou should now see `-rwsr-xr-x` permissions and ping should work for non-root users."
    }
  ],
  "metadata": {
    "topic": "File permission and ownership errors",
    "sub_angle": "setuid/setgid bits stripped after package upgrade breaking sudo helpers"
  }
}

hamma_devops_data.jsonl — Prompt/Completion format

Legacy format for pipelines that expect flat prompt/completion pairs:

{
  "prompt": ["After upgrading sudo package on Ubuntu 22.04..."],
  "completion": ["The setuid bit was stripped from `/usr/bin/ping`..."]
}

---

🚀 Loading the Dataset

from datasets import load_dataset

# Chat format (recommended for Gemma 4 / Unsloth)
ds = load_dataset("your-username/hamma-devops", data_files="hamma_devops_messages.jsonl", split="train")

# Prompt/completion format
ds = load_dataset("your-username/hamma-devops", data_files="hamma_devops_data.jsonl", split="train")

print(ds[0]["messages"])

---

🔧 Fine-tuning with Unsloth + Gemma 4

from unsloth import FastModel
from trl import SFTTrainer, SFTConfig
from datasets import load_dataset

model, tokenizer = FastModel.from_pretrained(
    model_name="unsloth/gemma-4-12b-it",
    max_seq_length=2048,
    load_in_4bit=True,
)

model = FastModel.get_peft_model(
    model,
    r=16,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
    lora_alpha=16,
    lora_dropout=0,
    bias="none",
)

ds = load_dataset("xayrullonematov/hamma-data", data_files="hamma_devops_messages.jsonl", split="train")

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=ds,
    args=SFTConfig(
        per_device_train_batch_size=2,
        gradient_accumulation_steps=4,
        warmup_steps=10,
        max_steps=300,
        learning_rate=2e-4,
        output_dir="hamma-gemma4",
    ),
)

trainer.train()

---

🧠 Input Style Variety

User messages intentionally cover multiple prompt styles to improve robustness:

Style	Example
Raw log dump	Multi-line journalctl / dmesg output with "what's wrong?"
Panicked one-liner	"prod is down, k8s pods crashlooping, help"
Calm technical question	"Why does PVC stay Pending when StorageClass exists?"
Direct command request	"give me the iptables rule to allow forwarding from br-abc to eth0"
Config paste	nginx.conf snippet followed by "this isn't working, why?"
Error message paste	Raw CLI error with no other context

---

⚙️ Generation Details

Parameter	Value
Generator model	Claude Sonnet (AWS Bedrock)
Generation method	Forced tool-use (structured JSON output)
Quality filters	Min 40 words, fenced code block required, no placeholder text
Deduplication	SHA1 hash of user_input + ai_response
Sub-angles per topic	~7
Raw pairs per sub-angle	10

The generator script uses forced tool-use (not free-form generation) to guarantee structured output, and applies quality filters that reject any response without a runnable code block. Zero duplicate pairs exist in the final dataset.

---

📄 License

MIT — free to use, modify, and redistribute for any purpose.

---

🔗 Links

• HAMMA Project: https://github.com/xayrullonematov/hamma
• Gemma 4 Good Hackathon: https://www.kaggle.com/competitions/gemma-4-good
• Unsloth: https://github.com/unslothai/unsloth