dgpl/dgpl-linux-assistant-0.8b

🐧 DGPL Linux Assistant 0.8B — Edge DevSecOps & Terminal Assistant

Specialized Lightweight Research Preview (800M Parameter Edge Checkpoint)

Developed & Released by the AI Systems Research Team at Durbhasi Gurukulam Private Limited (DGPL)
Jaipur, Rajasthan, Bharat (India) · support@durbhasigurukulam.com

🔬 RESEARCH PREVIEW & SCOPE ADVISORY

• Specialized Edge Target: DGPL Linux Assistant 0.8B is a domain-specialized, ultra-compact (800M parameter) research checkpoint fine-tuned on top of Qwen3.5-0.8B for single-turn Linux CLI generation, non-interactive diagnostic scripts, and edge devices with <1GB RAM.
• What it IS: A lightweight, sub-second terminal autocompletion and command synthesis helper for edge nodes, Raspberry Pis, and local developer workstations.
• What it IS NOT: It is NOT a full-scale conversational model, general-purpose LLM, or replacement for large reasoning models (such as LLaMA-70B or Claude/GPT-4). Due to its 0.8B scale, complex multi-hop causal reasoning or unformatted zero-shot queries outside system administration will yield degraded coherence.
• Human-in-the-Loop Requirement: Always inspect and verify generated shell commands before executing them in production environments.

---

⚡ Key Highlights & Target Use Cases

• Ultra-Compact Footprint: Runs comfortably in standard CPU containers and edge devices with <1.0 GB RAM.
• Non-Interactive Default: Emits non-blocking batch commands (top -b -n 1, ping -c 4, journalctl --no-pager, systemctl --no-pager) to prevent terminal hangs.
• Multimodal Visual Diagnostics: Includes optional vision projector weights (mmproj_linux_copilot_f16.gguf) for terminal screenshot and OCR triage.
• Sub-Second Execution: Direct Fast Mode generates concise command sequences in <0.35s on GPU / 1.5s–3.0s on CPU.

---

🚀 Recommended Usage & Prompt Formatting

Because this model has 800M parameters, strict ChatML template compliance and low temperature ($\le 0.2$) are mandatory for stable output.

1. Using with Ollama (Recommended)

# Pull and run the official 4-bit edge build
ollama run dgpl/dgpl-linux-assistant:latest

For Best CLI Results (Instant Shell Completion Mode):

Create a Modelfile.cli to enforce non-interactive formatting and prevent reasoning hallucinations:

FROM dgpl/dgpl-linux-assistant:latest

PARAMETER temperature 0.1
PARAMETER top_p 0.85
PARAMETER num_predict 128
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|im_start|>"
PARAMETER stop "```\n\n"

SYSTEM """You are DGPL Linux Assistant 0.8B, a specialized Linux Terminal helper developed by Durbhasi Gurukulam Private Limited (DGPL).
Given a user query, output the exact non-interactive Linux bash command in a markdown code block immediately without conversational filler."""

ollama create dgpl-cli -f Modelfile.cli
ollama run dgpl-cli "Find all files over 500MB modified in the last 7 days"

2. Using with Hugging Face Transformers (Python)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "dgpl/dgpl-linux-assistant-0.8b"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    device_map="auto"
)

system_prompt = (
    "You are DGPL Linux Assistant 0.8B, a specialized Linux Terminal helper by DGPL. "
    "Output the exact non-interactive Linux bash command in markdown code blocks."
)

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "Check open TCP listening sockets and their process names"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=128,
    temperature=0.1,
    top_p=0.85,
    do_sample=False
)

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

---

📊 Evaluation Summary & Technical Scope

In an internal synthetic domain evaluation covering 12 targeted system administration scenarios (socket probes, process monitoring, file discovery, systemd units, capabilities inspection):

• Structured Domain Accuracy: 11/12 (91.7%) on targeted Linux command syntax when prompted with appropriate system directives.
• Latency: Average ~16s on CPU (single-core) / <0.35s on NVIDIA RTX/Tesla GPUs.
• Evaluation Notice: These metrics reflect closed-domain single-command generation tests. Real-world performance on complex sysadmin scenarios or ambiguous natural language queries will vary and requires explicit prompt framing.

---

🛠️ Common Troubleshooting

Issue	Cause	Fix
Repetitive or runaway text	Temperature too high ($>0.5$) or missing stop tokens	Set temperature=0.1 and add `<
Conversational rambling instead of commands	Missing system prompt	Include the recommended SYSTEM prompt instructing direct command output.
Slow generation on CPU	Thread oversubscription	In Ollama / llama.cpp, pin --threads to physical CPU cores (e.g. 4 threads).

---

📦 Model Artifacts & Quantizations

File	Format	Size	Target Environment
linux_copilot_q4_k_m.gguf	Q4_K_M (Default)	529 MB	Edge devices, Raspberry Pi, Low-RAM VMs
linux_copilot_q5_k_m.gguf	Q5_K_M	578 MB	Lightweight Dev Servers
linux_copilot_q6_k.gguf	Q6_K	630 MB	Standard Workstations
linux_copilot_q8_0.gguf	Q8_0	812 MB	High-Precision Host Auditing
linux_copilot_f16.gguf	FP16 Master	1.51 GB	GPU Inference & Conversions
mmproj_linux_copilot_f16.gguf	Vision Projector	205 MB	Terminal Screenshot OCR Triage

---

🏢 Corporate Entity & Attribution

• Company Legal Name: Durbhasi Gurukulam Private Limited (DGPL)
• Registered Office: 2., P. NO. 18, Radha Vihar, Manchwa, Jaipur, Rajasthan 303706
• Official Inquiries: support@durbhasigurukulam.com
• License: GNU General Public License v3.0 (GPL-3.0)
• Base Architecture: Fine-tuned checkpoint derived from Qwen/Qwen3.5-0.8B.