---
license: apache-2.0
pipeline_tag: text-generation
library_name: transformers
tags:
  - mistral
  - merge
  - lora
  - fine-tuned
  - qlora
  - chat
  - reasoning
base_model:
  - HuggingFaceH4/zephyr-7b-beta
  - teknium/OpenHermes-2.5-Mistral-7B
  - cognitivecomputations/Dolphin-2.6-mistral-7b-dpo
---

# 🧟 Frankenstein 2.0

A custom AI assistant stitched together from three Mistral-7B models, 
then fine-tuned — built entirely on Kaggle.

## 🧬 What Is This?

Frankenstein 2.0 is a **merged model** combining the strengths of three 
open-source models, then fine-tuned with QLoRA on coding + general 
instruction data.

| Component | Contribution |
|---|---|
| Zephyr-7B-beta | Instruction following + structure |
| OpenHermes-2.5 | Warmth + conversational tone |
| Dolphin-2.6-dpo | Obedience + helpfulness |

## 🎓 Training Details

- **Base:** Merge of 3× Mistral-7B variants
- **Fine-tuning:** QLoRA (4-bit NF4)
- **LoRA:** rank 16, alpha 32
- **Data:** 3000 general + 3000 coding examples
- **Hardware:** 2× NVIDIA Tesla T4 (Kaggle)


## 💡 Capabilities

✅ Python coding & debugging  
✅ Explaining complex topics simply  
✅ Step-by-step reasoning (with `<think>` tags)  
✅ Document Q&A (RAG-ready)  
✅ Tool use (web search, calculator)

## 🚀 How To Use

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

model = AutoModelForCausalLM.from_pretrained(
    "Questionmarkboy/frankenstein-2.0",
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Questionmarkboy/frankenstein-2.0")

messages = [{"role": "user", "content": "Explain blockchain to a 10-year-old"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True,
    return_tensors="pt").to(model.device)
output = model.generate(inputs, max_new_tokens=200)
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

## ⚠️ Limitations

- 7B model — may struggle with very long code generation (e.g. full HTML apps)
- Knowledge limited to training data
- Best for English tasks

## 🙏 Credits

- Base models by [HuggingFaceH4](https://huggingface.co/HuggingFaceH4), 
  [Teknium](https://huggingface.co/teknium), and 
  [Cognitive Computations](https://huggingface.co/cognitivecomputations)
- Built with Kaggle

## 📜 License

Apache 2.0 (inherited from Mistral base). Please credit the original 
base models when using this.

---

*It's alive! 🧟⚡*