How facebookresearch/advprompter works
Project Type & Purpose
• This is a research/ML experimentation framework (“AdvPrompter”) for generating adversarial prompts (suffixes) to probe or attack large language models (LLMs).
• It uses Hydra for flexible configuration, supports training and evaluation pipelines, and optionally logs to Weights & Biases.
Main Components
a. Configuration Manager (Hydra)
b. Orchestrator (main.py)
c. Data Pipeline
· Dataset Loader (CSV, harmful_behaviors, prefixes)
· Suffix Dataset Writer
d. AdvPrompter Module
· Sequence Optimizer (advprompteropt.py & sequence.py)
· LoRA Wrappers (for finetuning)
e. Target LLM Module (llm.py)
· Model Interface (HuggingFace, PEFT checkpointing)
f. Evaluation Module
· Metrics & Loss Computation
· Suffix‐level evaluation (eval, eval_suffix_dataset flows)
g. Training Module
· Training Loop (including intermittent evaluation & checkpointing)
h. Logging & Experiment Tracking
· File System (exp/... directories)
· Optional: Weights & Biases API
i. Utilities (utils.py)
Relationships & Data Flow
• Orchestrator → Data Pipeline for loading train/test CSVs
• Data Pipeline → AdvPrompter: provides prompts to optimize
• AdvPrompter → Sequence Optimizer: generates candidate suffixes
• Sequence Optimizer ↔ Target LLM: queries model for outputs & gradients
• Target LLM → Evaluation Module: returns scores/loss
• Evaluation Module → Logging (WandB/file system)
• Training Loop: connects AdvPrompter ↔ Target LLM ↔ Evaluation ↔ Checkpointing
• Evaluation-only flow: skip Training Loop, produce suffix_dataset → Evaluation Module → Logging
Architectural Patterns & Design Principles
• Modular design: each concern (config, data, model, eval, logging) is a separate package/file.
• Plug-and-play with Hydra: drop-in new LLMs or prompters via YAML.
• Separation of training vs. evaluation workflows configured by Hydra.
• Use of PEFT/LoRA for efficient finetuning.
• File‐based experiment tracking (exp/…), with optional external service (WandB).
Technologies & Frameworks
• Python 3.11, Hydra (config management)
• HuggingFace Transformers + PEFT/LoRA
• Pandas/CSV for data loading
• PyTorch for model training
• WandB (optional)
• Singularity/Conda for reproducible environments