How hari-goml/retrace works
Retrace prepares a structured fact knowledge base, trains a baseline model, selectively erases requested facts with a stackable LoRA adapter, verifies forgetting and retention, and produces a report. Users can run these stages through the CLI or interact with the serving UI, which supports live adapter switching and erasure workflows. The graph preserves documented preparation capabilities such as paraphrase, probe, control-fact, and neighbor generation. Some source was not sampled; edges are limited to relationships visible in excerpts, the SOURCE INDEX, or README statements.
Main entry points: build.py Knowledge builder · cli.py CLI commands · harness.py Retention verification · engine.py Serving engine
Components 29
Knowledge preparation
- Fact schema
retrace/knowledge/schema.py - Fact dataset
retrace/knowledge/dataset.py - Knowledge builder
retrace/knowledge/build.py - Fact paraphrases
retrace/knowledge/paraphrase.py - Probe generation
retrace/knowledge/probes.py - Control facts
retrace/knowledge/controls.py - Neighbor graph
retrace/knowledge/neighbors.py - Fact resolver
retrace/knowledge/resolver.py
Baseline training
- Training builder
retrace/training/build.py - Training data
retrace/training/data.py - Baseline fine-tuning
retrace/training/finetune.py - Baseline quality gate
retrace/training/gate.py
Selective erasure
- Erasure request
retrace/erasure/request.py - Erasure builder
retrace/erasure/build.py - LoRA unlearning
retrace/erasure/unlearn.py
Verification and reporting
- Retention verification
retrace/verification/harness.py - Adversarial probes
retrace/verification/adversarial.py - Verification metrics
retrace/verification/metrics.py - Erasure report
retrace/reporting/generate.py - Markdown and HTML
retrace/reporting/markdown_html.py
User runtime
- CLI commands
retrace/cli.py - Stage configuration
retrace/config.py - Model inference
retrace/modeling.py - Answer scoring
retrace/scoring.py - Interactive UI
retrace/serving/app.py - Serving engine
retrace/serving/engine.py - Pipeline artifacts
retrace/serialize.py - Domain errors
retrace/exceptions.py
Other components
- Engineer
Connections 40
- Engineer → CLI commands: runs commands
README.md - CLI commands → Knowledge builder: starts preparation
retrace/cli.py - CLI commands → Training builder: starts training
retrace/cli.py - CLI commands → Erasure builder: starts erasure
retrace/cli.py - CLI commands → Retention verification: starts verification
retrace/cli.py - CLI commands → Erasure report: starts reporting
retrace/cli.py - Knowledge builder → Fact dataset: loads facts
retrace/knowledge/build.py - Fact dataset → Fact schema: validates records
retrace/knowledge/dataset.py - Knowledge builder → Fact paraphrases: generates paraphrases
retrace/knowledge/build.py - Knowledge builder → Probe generation: generates probes
retrace/knowledge/build.py - Knowledge builder → Control facts: builds controls
retrace/knowledge/build.py - Knowledge builder → Neighbor graph: builds neighbors
retrace/knowledge/build.py - Training builder → Baseline fine-tuning: trains baseline
retrace/training/build.py - Training builder → Baseline quality gate: checks baseline
retrace/training/build.py - Baseline fine-tuning → Training data: loads training data
retrace/training/finetune.py - Baseline quality gate → Model inference: scores probes
retrace/training/gate.py - Baseline quality gate → Answer scoring: aggregates answers
retrace/training/gate.py - Erasure builder → Erasure request: resolves selection
retrace/erasure/build.py - Erasure builder → LoRA unlearning: trains adapter
retrace/erasure/build.py - Erasure request → Fact dataset: selects facts
retrace/erasure/request.py - Erasure request → Neighbor graph: selects hard retains
retrace/erasure/request.py - LoRA unlearning → Training data: uses collator
retrace/erasure/unlearn.py - LoRA unlearning → Model inference: uses inference wrapper
retrace/erasure/unlearn.py - Retention verification → Model inference: evaluates outputs
retrace/verification/harness.py - Retention verification → Adversarial probes: uses adversarial probes
retrace/verification/harness.py - Retention verification → Verification metrics: computes metrics
retrace/verification/harness.py - Retention verification → Answer scoring: scores answers
retrace/verification/harness.py - Erasure report → Markdown and HTML: renders report
retrace/reporting/generate.py - Interactive UI → Serving engine: uses engine
retrace/serving/app.py - Serving engine → Fact resolver: resolves queries
retrace/serving/engine.py - Serving engine → Model inference: generates answers
retrace/serving/engine.py - Serving engine → Erasure request: resolves erasure
retrace/serving/engine.py - Serving engine → Retention verification: runs verification
retrace/serving/engine.py - Serving engine → Erasure report: loads reports
retrace/serving/engine.py - Knowledge builder → Pipeline artifacts: writes artifacts
retrace/knowledge/build.py - Training builder → Pipeline artifacts: writes artifacts
retrace/training/build.py - Erasure builder → Pipeline artifacts: writes artifacts
retrace/erasure/build.py - Retention verification → Pipeline artifacts: writes results
retrace/verification/harness.py - Erasure report → Pipeline artifacts: reads artifacts
retrace/reporting/generate.py - CLI commands → Stage configuration: sets options
retrace/cli.py