guidance-dynamic-prompt-lora

Guidance Dynamic Prompt Generator (QLoRA Fine-tuned)

This model generates dynamic LLM-control prompts based on persona sliders and multiple custom guidance instructions.
It is fine-tuned using QLoRA on top of Mistral-7B-Instruct, enabling lightweight training while preserving strong reasoning capabilities.

The model takes inputs such as:

• Formality (0–5)
• Empathy (0–5)
• Conciseness (0–5)
• Enthusiasm (0–5)
• Multiple custom-gudance blocks
  • category
  • title
  • description

and produces a fully structured prompt that can be fed into any LLM (GPT, Llama, Mistral, Gemini, Claude, etc.) to control tone, style, and behavioral constraints.

---

Model Details

Model Description

• Developer: Vishnu Maddukuri
• Fine-tuned by: Vishnu
• Base Model: mistralai/Mistral-7B-Instruct-v0.3
• Parameter Efficient Method: QLoRA (4-bit)
• Model type: Decoder-only causal language model
• Language: English
• License: Apache 2.0
• Intended for: Dynamic prompt construction based on structured guidance inputs

Model Sources

• Base model repository: https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3
• LoRA Adapter repo: https://huggingface.co/vishnu/guidance-dynamic-prompt-lora

---

Uses

Direct Use

Use the model to generate high-quality system prompts conditioned on:

• tone parameters (formality, empathy, etc.)
• personalization sliders
• contextual constraints
• multiple custom guidance rules

The output prompt can then be passed into any LLM to influence its behavior.

Downstream Use

• Customer support agents
• Chatbot tone-control frameworks
• LLM assistant configuration
• Multi-persona systems
• Enterprise support/ticketing AI
• RAG systems requiring personality modulation

Out-of-Scope Use

• Generating factual knowledge
• Replacing the base model
• Real-time safety-critical decision making
• Use cases requiring unbiased raw model reasoning

---

Bias, Risks, and Limitations

Known Limitations

• The model may overfit if trained on small datasets.
• Model is not intended to generate user-facing content directly — only meta-prompts.
• May generate verbose or overly structured prompts depending on training distribution.

Risks

• If custom guidance instructions contain harmful strategies, the model will generate unsafe prompt text (the downstream LLM should handle safety).
• Overly detailed persona settings may cause deterministic outputs.

Recommendations

• Always validate the generated prompt before sending it to another LLM.
• Use additional safety layers (OpenAI/Anthropic safety systems).
• Additional fine-tuning with more diverse guidance examples will improve generalization.

---

How to Get Started

Load Adapter + Base Model

from peft import PeftModel
from transformers import AutoTokenizer, AutoModelForCausalLM

base_model = "mistralai/Mistral-7B-Instruct-v0.3"
adapter = "vishnu/guidance-dynamic-prompt-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

prompt = "Formality: 2, Empathy: 4, Conciseness: 3, Enthusiasm: 5..."

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=400)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))