Personalized LLM via Chameleon approach
🧠 Chameleon + PriME: Personalized & Privacy-Preserving LLM Optimization
This repository contains code, configuration, and experimental framework for a three-month research project that integrates Chameleon-style embedding personalization and PriME-style evolutionary model merging to achieve efficient, private, and high-performance LLM-based recommendation systems.
This project aims to:
- Personalize LLMs using user-specific directions in embedding space (via Chameleon)
- Optimize merged models via evolutionary algorithms (PriME)
- Support multi-user inference with limited data and strong privacy constraints
chameleon_prime_personalization/ ├── configs/ # YAML/JSON configs for models, training, evaluation ├── data/ │ ├── raw/ # Raw datasets (LaMP-2, LaMP-3, etc.) │ └── processed/ # Tokenized or embedded versions ├── models/ │ ├── base_model/ # Base LLM (e.g., LLaMA 2 7B or Mistral) │ └── user_adapters/ # LoRA/PEFT modules for users ├── notebooks/ # Experiment tracking or visualization notebooks ├── scripts/ │ ├── download_models.py │ ├── download_datasets.py │ └── setup_environment.sh └── README.md
bash scripts/setup_environment.sh
source env/bin/activate
2. Download base models
python scripts/download_models.py --model mistralai/Mistral-7B-v0.1
3. Download datasets
python scripts/download_datasets.py --dataset lamp2
🧩 Method Summary
Embedding Personalization (Chameleon)
SVD-based decomposition of user history embeddings
Extracts user-specific direction vs general direction
Adjust embeddings by shifting along user-personalized vector
Evolutionary Merge (PriME)
Each user gets LoRA/IA3 fine-tuned PEFT modules
Cosine similarity used to identify similar shared users
Evolutionary strategy (CMA-ES or NSGA-II) to merge modules
Objective: maximize user utility (F1, ROUGE), minimize privacy leakage
📊 Evaluation Metrics
ROUGE / F1 / BLEU (text output quality)
Cosine similarity / KL divergence (privacy leakage)
Model diversity and memorization metrics (advanced)
## Strict Compliance Prompt Pack (LaMP-2)
**目的**: 出力を `Answer: <TAG>` の単一行に強制し、形式準拠率 ≥ 0.95 を安定達成。
### ✅ 実測
- 形式準拠率: **98.0%**（目標 95% 超）
- テスト規模: **50 samples**
- 出力形式: 単一行 `Answer: <TAG>`
- デコード制約: `temperature=0, top_p=0, max_tokens=8, stop=["\n"]`
- 厳格検証パターン: `^Answer:\s*([A-Za-z0-9_\- ]+)\s*$`
### 🔧 プロンプト
**SYSTEM**You are a strict single-line tag classifier.
RULES (絶対遵守):
- Output EXACTLY one line: Answer:
- No explanations, no extra words, no punctuation after , no emojis.
- NO NEWLINES. Output must be a single line only.
- Choose ONE best tag from the allowed list, case-sensitive.
- If uncertain, still pick the single best tag.
FORBIDDEN: • Multiple lines or trailing spaces • Any text before/after Answer:
Allowed tags: {{ALLOWED_TAGS}} Required output format: Answer:
**USER**
Task
Classify the following movie description into exactly one tag from the allowed list.
Description
{{QUESTION}}
User Profile (optional)
{{USER_PROFILE}}
Allowed tags (pick ONE, case-sensitive)
{{ALLOWED_TAGS}}
Output constraints (絶対遵守) • Single line only. • EXACT string format: Answer: • Nothing else before or after. • No newline characters.
Your response
Answer:
### 🧪 使い方
```bash
# 基本テスト
python run_strict_compliance_test.py --samples 10
# 実データでのテスト
python run_strict_compliance_test.py --data path/to/lamp2_test.jsonl
# カスタムプロンプト
python run_strict_compliance_test.py \
--system-prompt prompts/lamp2_system_strict.txt \
--user-template prompts/lamp2_user_template_strict.txt \
--target-compliance 0.95
• 無関係出力の排除（例: "Source: …" 等） • LaMP-2向けに最適化された単一タグ選択 • 決定論的デコード設定で再現性担保 • ≥95% 準拠で評価の信頼性向上
📄 License This project is licensed under the Apache License 2.0. See LICENSE for details.
📚 References Chameleon: Personalized Prompt Editing for Large Language Models
PriME: Personalized Model Merging via Evolutionary Search
✍️ Acknowledgments This repository was developed for academic research on privacy-preserving personalization using large language models, with support for datasets such as LaMP-2 and LaMP-3, and tested on NVIDIA A100×2 environment.