Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,105 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- lora
|
| 5 |
+
- fine-tuning
|
| 6 |
+
- training
|
| 7 |
+
- identity-replacement
|
| 8 |
+
- catastrophic-forgetting
|
| 9 |
+
- progressive-merging
|
| 10 |
+
language:
|
| 11 |
+
- en
|
| 12 |
+
library_name: transformers
|
| 13 |
+
pipeline_tag: text-generation
|
| 14 |
+
---
|
| 15 |
+
|
| 16 |
+
# 🧟 Body Snatching: Progressive LoRA Merging (PLM)
|
| 17 |
+
|
| 18 |
+
**Complete model identity replacement using only LoRA-level resources.**
|
| 19 |
+
|
| 20 |
+
> *"What if catastrophic forgetting is a feature, not a bug?"*
|
| 21 |
+
|
| 22 |
+
## 🔥 What is this?
|
| 23 |
+
|
| 24 |
+
**Progressive LoRA Merging (PLM)** is a training methodology that lets you completely replace a model's identity—its personality, reasoning patterns, and learned behaviors—while keeping the architecture intact.
|
| 25 |
+
|
| 26 |
+
Think of it as **body snatching** for LLMs:
|
| 27 |
+
- The **body** (architecture, tokenizer, attention mechanisms) stays
|
| 28 |
+
- The **soul** (personality, knowledge, behavior) gets replaced
|
| 29 |
+
|
| 30 |
+
After enough cycles, you don't have "Qwen fine-tuned for X". You have **a completely different model** that happens to use Qwen's skeleton.
|
| 31 |
+
|
| 32 |
+
## 💡 The Key Insight
|
| 33 |
+
|
| 34 |
+
Everyone treats **catastrophic forgetting** as a problem to avoid.
|
| 35 |
+
|
| 36 |
+
We treat it as **the goal**.
|
| 37 |
+
|
| 38 |
+
## 🔄 How It Works
|
| 39 |
+
|
| 40 |
+
```
|
| 41 |
+
Cycle 1: Base Model → Train LoRA → Merge → New Base₁
|
| 42 |
+
Cycle 2: New Base₁ → Train LoRA → Merge → New Base₂
|
| 43 |
+
...
|
| 44 |
+
Cycle N: New Base_N = Completely Different Model
|
| 45 |
+
```
|
| 46 |
+
|
| 47 |
+
Each cycle:
|
| 48 |
+
1. **Train** a small LoRA adapter (~0.1% of parameters)
|
| 49 |
+
2. **Merge** it permanently into the base weights (in BF16, not 4-bit!)
|
| 50 |
+
3. **Fresh LoRA** for the next cycle
|
| 51 |
+
4. **Repeat** until original identity is gone
|
| 52 |
+
|
| 53 |
+
## 📊 Results
|
| 54 |
+
|
| 55 |
+
| Cycles | Similarity to Original | Target Identity Match |
|
| 56 |
+
|--------|------------------------|----------------------|
|
| 57 |
+
| 0 | 100% | 0% |
|
| 58 |
+
| 25 | 64% | 41% |
|
| 59 |
+
| 50 | 28% | 73% |
|
| 60 |
+
| 100 | **7%** | **94%** |
|
| 61 |
+
|
| 62 |
+
After 100 cycles, the model is **93% your data, 7% original**.
|
| 63 |
+
|
| 64 |
+
## 💰 Resource Comparison
|
| 65 |
+
|
| 66 |
+
| Method | Hardware | Time | Cost | Result |
|
| 67 |
+
|--------|----------|------|------|--------|
|
| 68 |
+
| Full Fine-tune | 4-8x A100 | Weeks | $10,000+ | Complete replacement |
|
| 69 |
+
| Single LoRA | 1x 24GB | Hours | $10 | Surface adaptation |
|
| 70 |
+
| **PLM (Ours)** | 1x 24GB | Days | $100-500 | **Complete replacement** |
|
| 71 |
+
|
| 72 |
+
## 🚀 Quick Start
|
| 73 |
+
|
| 74 |
+
```bash
|
| 75 |
+
pip install torch transformers peft bitsandbytes datasets
|
| 76 |
+
|
| 77 |
+
python plm.py --base-model Qwen/Qwen3-1.7B --dataset data.jsonl --cycles 100
|
| 78 |
+
```
|
| 79 |
+
|
| 80 |
+
## 📖 Citation
|
| 81 |
+
|
| 82 |
+
```bibtex
|
| 83 |
+
@article{drissi2024bodysnatching,
|
| 84 |
+
title={Body Snatching: Complete Model Identity Replacement via Progressive LoRA Merging},
|
| 85 |
+
author={Drissi, Ouissam Said},
|
| 86 |
+
year={2024},
|
| 87 |
+
url={https://github.com/antibitcoin/progressive-lora-merging}
|
| 88 |
+
}
|
| 89 |
+
```
|
| 90 |
+
|
| 91 |
+
## 🔗 Links
|
| 92 |
+
|
| 93 |
+
- **GitHub**: [antibitcoin/progressive-lora-merging](https://github.com/antibitcoin/progressive-lora-merging)
|
| 94 |
+
- **Paper**: [PAPER.md](https://github.com/antibitcoin/progressive-lora-merging/blob/main/PAPER.md)
|
| 95 |
+
- **Related Work**: [ASRL Paper (IJSET 2025)](https://www.ijset.in/wp-content/uploads/IJSET_V13_issue5_102.pdf)
|
| 96 |
+
|
| 97 |
+
## 👤 Author
|
| 98 |
+
|
| 99 |
+
**Ouissam Said Drissi**
|
| 100 |
+
- Email: wissam.idrissi@gmail.com
|
| 101 |
+
- Independent Researcher, Morocco
|
| 102 |
+
|
| 103 |
+
---
|
| 104 |
+
|
| 105 |
+
*"You're not fine-tuning a model. You're growing a new one inside its skeleton."*
|