Skip to content

PHASE2_PLAN.md

Latest commit

 

History

History
52 lines (41 loc) · 1.93 KB

PHASE2_PLAN.md

File metadata and controls

52 lines (41 loc) · 1.93 KB

PHASE 2: FEATURE ENGINEERING - PLAN

Goal: Convert raw text prompts into numerical vectors for machine learning models.

🛠 Feature Sets to Extract

1. Lexical Features (Statistical)

Simple but effective features based on text statistics.

  • Length: Character count, word count
  • Capitalization: Ratio of uppercase letters (shouting/imperatives)
  • Punctuation: Count of !, ?, ", {, } (often used in code injections)
  • Special Characters: Count of @, #, $, %
  • Keyword Presence: Binary flags for suspicious words:
    • ignore, override, system, developer, model, prior
    • jailbreak, mode, unfiltered, dan

2. Structural Features (Pattern-based)

Detecting the structure of attacks.

  • Role Play Indicators: Matches for You are, Act as, Pretend
  • Chat Template Injection: Matches for User:, Assistant:, System:
  • Code Injection: Presence of code blocks or programming syntax (def , import , print()

3. Semantic Features (Embeddings)

Capturing the meaning of the text.

  • Sentence-BERT Embeddings: Use all-MiniLM-L6-v2 (fast & effective)
    • Output: 384-dimensional vector for each prompt

📝 Execution Plan

  1. Create src/feature_extraction.py:

    • Class FeatureExtractor
    • Setup scikit-learn transformers and pipelines
    • Setup sentence-transformers model

  2. Process Data:

    • Load data/processed/prompts.csv
    • Apply extraction in batches (embeddings can be slow)
    • Combine all features into a single feature matrix ($X$)

  3. Save Features:

    • Save $X$ (features) and $y$ (labels) as .npz (NumPy arrays) or .parquet
    • data/processed/X_features.npz
    • data/processed/y_labels.npy

⏱ Estimated Time

  • Lexical/Structural: < 1 minute
  • Embeddings: ~15-30 minutes (on CPU) for 90k samples

🚀 Next Command

python src/feature_extraction.py