How vaniseth/ggas works
GGAS is a research system for detecting multi-turn jailbreak risk in language-model conversations and applying activation steering. Its principal workflow generates attack and benign transcripts, extracts per-turn activations and attention, trains a probe, computes cumulative graph-based risk, calibrates a threshold and steering strength, then evaluates defended behavior. Separate capabilities provide baseline guards, evaluation tables, and graph snapshots/visualization. The README is only a title, so workflow detail comes from sampled source and the source index; unsampled subsystem links are included only where that index documents references.
Main entry points: pipeline.py Pipeline stages · extract.py Extracted features · probe.py Risk probe · steer.py Steered target
Components 20
Experiment orchestration
- Pipeline stages
ggas/pipeline.py - Shared config
ggas/config.py - Dataset loaders
ggas/data_loaders.py - Attack harness
ggas/attack_harness.py - Conversation transcripts
ggas/pipeline.py
Risk signals
- Extracted features
ggas/extract.py - HF compatibility
ggas/hf_compat.py - Risk probe
ggas/probe.py
Detection and defense
- RDS detector
ggas/detect.py - Calibration
ggas/calibrate.py - Steered target
ggas/steer.py - GGAS responder
ggas/ggas_responder.py - Model swapping
ggas/swap_runnner.py
Evaluation and visualization
- Evaluation suite
ggas/eval_suite.py - Evaluation table
ggas/eval_table.py - LlamaGuard baseline
ggas/llamaguard_baseline.py - Graph snapshots
ggas/graphviz/state.py - Graph capture
ggas/graphviz/capture.py - Graph rendering
ggas/graphviz/render.py
Other components
- Researcher
Connections 40
- Researcher → Pipeline stages: runs stages
ggas/pipeline.py - Pipeline stages → Shared config: reads settings
ggas/pipeline.py - Pipeline stages → Dataset loaders: loads seeds
ggas/pipeline.py - Pipeline stages → Attack harness: runs conversations
ggas/pipeline.py - Attack harness → Conversation transcripts: produces
ggas/pipeline.py - Pipeline stages → Conversation transcripts: writes transcripts
ggas/pipeline.py - Pipeline stages → Extracted features: extracts features
ggas/pipeline.py - Extracted features → HF compatibility: normalizes model outputs
ggas/extract.py - Extracted features → Attack harness: uses target
ggas/extract.py - Pipeline stages → Risk probe: trains probe
ggas/pipeline.py - Risk probe → Extracted features: reads activations
ggas/probe.py - Pipeline stages → Calibration: calibrates defense
ggas/pipeline.py - Calibration → Risk probe: scores candidates
ggas/calibrate.py - Calibration → RDS detector: tunes threshold
ggas/calibrate.py - Calibration → Steered target: tunes strength
ggas/calibrate.py - Pipeline stages → Steered target: configures steering
ggas/pipeline.py - RDS detector → Risk probe: scores nodes
ggas/detect.py - RDS detector → Extracted features: uses attention graph
ggas/detect.py - Steered target → RDS detector: detects risk
ggas/steer.py - Steered target → Risk probe: uses probe
ggas/steer.py - Steered target → Extracted features: extracts current features
ggas/steer.py - Steered target → HF compatibility: normalizes model outputs
ggas/steer.py - GGAS responder → Attack harness: uses target
ggas/ggas_responder.py - GGAS responder → Extracted features: extracts features
ggas/ggas_responder.py - GGAS responder → RDS detector: detects risk
ggas/ggas_responder.py - GGAS responder → Risk probe: scores nodes
ggas/ggas_responder.py - Model swapping → Attack harness: uses attack harness
ggas/swap_runnner.py - Model swapping → Dataset loaders: loads datasets
ggas/swap_runnner.py - Pipeline stages → Evaluation suite: runs evaluation
ggas/pipeline.py - Evaluation suite → Risk probe: uses probe
ggas/eval_suite.py - Evaluation suite → RDS detector: scores turns
ggas/eval_suite.py - Evaluation suite → Extracted features: extracts features
ggas/eval_suite.py - Evaluation table → Attack harness: runs model phases
ggas/eval_table.py - Evaluation table → Dataset loaders: loads evaluation sets
ggas/eval_table.py - Evaluation table → LlamaGuard baseline: compares baseline
ggas/eval_table.py - Evaluation table → Steered target: evaluates defense
ggas/eval_table.py - Graph capture → Graph snapshots: builds snapshots
ggas/graphviz/capture.py - Graph snapshots → Extracted features: extracts snapshot features
ggas/graphviz/state.py - Graph snapshots → RDS detector: scores snapshot
ggas/graphviz/state.py - Graph snapshots → Risk probe: uses risk probe
ggas/graphviz/state.py