---
license: apache-2.0
language:
- en
base_model: Qwen/Qwen2.5-14B-Instruct
tags:
- pharmacovigilance
- gvp
- regulatory
- qwen2.5
- medical
- drug-safety
- instruction-tuning
- gguf
- safetensors
- llama.cpp
- ollama
pipeline_tag: text-generation
library_name: transformers
---

# 💊 Open-GVP-Qwen2.5-14B-Instruct

> Domain-adapted **Qwen2.5-14B** model specialized in **Good Pharmacovigilance Practices (GVP)** guidelines issued by the European Medicines Agency (EMA).

This repository contains both the **merged Safetensors** version and **GGUF quantized** versions of the model.

**Available formats:**
- Merged Safetensors (for Transformers / vLLM / etc.)
- GGUF: `BF16`, `Q8_0`, `Q6_K`

---

## 📖 Model Description

**Open-GVP-Qwen2.5-14B-Instruct** is a domain-specialized version of [Qwen/Qwen2.5-14B-Instruct](https://huggingface.co/Qwen/Qwen2.5-14B-Instruct), fine-tuned using **LoRA** on a curated dataset of approximately **15,000 high-quality question-answer pairs** derived from the official EMA Good Pharmacovigilance Practices (GVP) guidelines.

The model covers **all GVP Modules and related Addendums**, with particular strength in areas such as:

- ICSR collection, management, and submission (Module VI)
- Signal management (Module IX)
- Risk management systems (Module V)
- Periodic safety update reports (PSUR / PBRER)
- Pharmacovigilance system master file (PSMF)

> **Important**: This model performs best when used as part of a **RAG (Retrieval-Augmented Generation)** pipeline alongside the original GVP PDF documents, rather than as a standalone source of regulatory advice.

---

## 🗂️ Coverage

| Category | Details |
| :--- | :--- |
| **GVP Modules** | All Modules |
| **Addendum** | Included |
| **Training Data Size** | ~15,000 instruction-format Q&A pairs |
| **Primary Focus** | Regulatory interpretation & PV operations |

---

## 📦 Available Formats

| Format | Files | Best For |
| :--- | :--- | :--- |
| **Merged Safetensors** | `model-00001-of-00004.safetensors` (4 shards) + config | Transformers, vLLM, TGI, full-precision inference |
| **GGUF BF16** | `Open-GVP-Qwen2.5-14B-BF16.gguf` | Highest quality GGUF |
| **GGUF Q8_0** | `Open-GVP-Qwen2.5-14B-Q8_0.gguf` | Excellent quality / speed balance |
| **GGUF Q6_K** | `Open-GVP-Qwen2.5-14B-Q6_K.gguf` | Good quality with lower resource usage |

**Recommendation**:  
- Use **Safetensors** for maximum quality and flexibility.  
- Use **Q8_0** or **Q6_K** GGUF for local / CPU-friendly deployment.

---

## 🚀 Quick Start

### 1. Transformers (Safetensors)

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

model_name = "drvivekpoojary/Open-GVP-Qwen2.5-14B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [
    {"role": "system", "content": "You are an expert pharmacovigilance assistant specialized in EMA Good Pharmacovigilance Practices (GVP)."},
    {"role": "user", "content": "What is the definition of a serious adverse reaction according to GVP Module VI?"}
]

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

outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.2)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### 2. Ollama (GGUF)

```bash
# Recommended (Q8_0)
ollama run hf.co/drvivekpoojary/Open-GVP-Qwen2.5-14B-Instruct:Q8_0

# Alternative options
ollama run hf.co/drvivekpoojary/Open-GVP-Qwen2.5-14B-Instruct:Q6_K
ollama run hf.co/drvivekpoojary/Open-GVP-Qwen2.5-14B-Instruct:BF16
```

### 3. llama.cpp

```bash
./llama-cli \
  -m Open-GVP-Qwen2.5-14B-Q8_0.gguf \
  -p "What is the definition of a serious adverse reaction according to GVP Module VI?" \
  -n 512 \
  -c 4096 \
  --temp 0.2
```

### 4. Python (llama-cpp-python)

```python
from llama_cpp import Llama

llm = Llama(
    model_path="Open-GVP-Qwen2.5-14B-Q8_0.gguf",
    n_ctx=4096,
    n_gpu_layers=-1,      # set to 0 for pure CPU
    verbose=False
)

response = llm.create_chat_completion(
    messages=[
        {
            "role": "system",
            "content": "You are an expert pharmacovigilance assistant specialized in EMA Good Pharmacovigilance Practices (GVP)."
        },
        {
            "role": "user",
            "content": "What is the definition of a serious adverse reaction according to GVP Module VI?"
        }
    ],
    max_tokens=512,
    temperature=0.2
)

print(response["choices"][0]["message"]["content"])
```

---

## ✅ Recommended Use Cases

| Use Case | Description |
| :--- | :--- |
| **GVP Knowledge Assistant** | Answer questions on GVP modules, definitions, and requirements |
| **PV Staff Training & Onboarding** | Support training of new pharmacovigilance team members |
| **RAG Pipeline** | Use as the generator together with official GVP PDFs |
| **Internal Regulatory Chatbot** | Backend for company-internal PV compliance assistants |
| **Offline / Air-gapped Environments** | Run completely locally without internet access |
| **Edge & Low-Resource Deployment** | Suitable for laptops and workstations (especially GGUF versions) |

### ❌ Not Recommended For

- Standalone regulatory decision-making
- High-stakes compliance or submission decisions without human review
- Replacing qualified pharmacovigilance professionals
- Use outside the scope of EMA GVP guidelines
- Generating content for regulatory submissions without expert verification

---

## 🔧 Training Details

| Property | Value |
| :--- | :--- |
| **Base Model** | Qwen/Qwen2.5-14B-Instruct |
| **Fine-tuning Method** | QLoRA |
| **LoRA Rank** | 32 |
| **LoRA Alpha** | 64 |
| **LoRA Dropout** | 0.05 |
| **Training Data** | ~15,000 GVP Q&A pairs (all modules) |
| **Epochs** | 4 |
| **Context Length** | 768 |
| **Precision** | bfloat16 |
| **Framework** | LlamaFactory |
| **Hardware** | NVIDIA GPU (16 GB VRAM) |

---

## ⚠️ Disclaimer

- This model is intended for **research, educational, and internal professional support purposes only**.

- It does **not** constitute regulatory advice. All outputs should be carefully reviewed by qualified pharmacovigilance professionals before being used in any compliance, case processing, reporting, or decision-making context.

- The model may produce incomplete, outdated, or inaccurate responses, particularly on complex or nuanced regulatory questions. The author assumes no liability for any decisions made based on the model’s outputs.

- **No In-Training Evaluation:** Evaluation loss (`eval_loss`) was not computed during training, and per-epoch checkpoints were not preserved.
---

## 👤 Author

**Dr. Vivek Poojary**  

---

## 📄 License

Apache License 2.0