SoftwareArchitecture-Instruct-v1

Domain: Software Architecture (for technical professionals)
Type: Instruction-tuned LLM
Base: LiquidAI/LFM2-1.2B (1.2 B parameter hybrid edge-optimized model) :contentReference[oaicite:1]{index=1}
Fine-tuned on: ajibawa-2023/Software-Architecture dataset
Author: Mohamed Yasser (yasserrmd)

---

​ Model Description

SoftwareArchitecture-Instruct-v1 is an instruction-tuned adaptation of LiquidAI’s lightweight and efficient LFM2-1.2B model. It’s specifically tailored to deliver high-quality, accurate, and technically rich responses to questions about software architecture—designed with engineers and architects in mind.

The base model, LFM2-1.2B, features a 16-layer hybrid design (10 convolutional + 6 grouped query attention layers), supports a 32,768 token context, and offers fast inference on CPU, GPU, and NPU platforms—ideal for both cloud and edge deployments :contentReference[oaicite:2]{index=2}.

---

​ Benchmark Summary

We performed a 50-prompt benchmark across diverse software architecture topics:

Metric	Value
Average Words per Response	~144
Median Words per Response	~139
Min / Max Words per Response	47 / 224
Avg Sentences per Output	~8.6
Lexical Diversity (TTR)	~0.73
Readability Complexity	High (professional-level)
Accuracy (topic keyword coverage)	Majority ≥ 60%
Off-topic Responses	None detected

Interpretation:

• Responses are substantive and domain-appropriate for technical audiences.
• Coverage is strong—while a few answers could benefit from including extra keywords, the core technical content is accurate.
• Readability intentionally leans into complexity, aligning with expert users.

---

​ Intended Use

• Ideal for: Software architects, system designers, engineering leads, and experienced developers seeking architecture guidance.
• Use cases include:
  • Exploring architectural patterns (e.g., CQRS, Saga, API Gateway).
  • Drafting design docs and decision rationale.
  • Architectural interview prep and system design walkthroughs.

Not intended for:

• Non-technical or general-purpose Q&A.
• In-depth code generation or debugging without architectural focus.

---

​ Usage Example

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "yasserrmd/SoftwareArchitecture-Instruct-v1"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")

messages = [
    {"role": "user", "content": "Explain the Saga pattern with orchestration and choreography."}
]

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

outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.3,
    repetition_penalty=1.05
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Training Details

• Base model: LiquidAI/LFM2-1.2B, optimized for edge/CPU inference ([ai.plainenglish.io][1], [generativeai.pub][2], [AI Models][3], [marktechpost.com][4], [Hugging Face][5])
• Dataset: ajibawa‑2023/Software‑Architecture
• Fine-tuning: Supervised instruction tuning
• (Optionally include parameters if available—epochs, LR, hardware used)

---

Limitations

• Answer length is capped by max_new_tokens. Some responses may truncate mid-explanation—raising this limit improves completeness.
• Keyword coverage is strong but not exhaustive. A few responses could benefit from enriching with additional terms.
• Not a replacement for expert-reviewed architectural validation—use as a support tool, not the final authority.

---

License

• Base model license: LFM Open License v1.0 ([Hugging Face][6])
• Dataset license: (Insert dataset license if known)

---

Author

Mohamed Yasser – Hugging Face profile