What is the Realignment Meta-Framework?

A shared 7-stage pipeline with quality thresholds and optimization guidance used across realignment workflows.

How are quality thresholds enforced?

Metrics like task accuracy, capability retention, data quality, and evaluation coverage have minimum/target/critical levels; automatic rollback triggers if thresholds fail.

How does rollback work across stages?

Baseline evaluation is run before/after each stage; any drop >5% triggers automatic rollback and cross-domain contamination checks.

realign-meta-framework

npx machina-cli add skill akaszubski/autonomous-dev/realign-meta-framework --openclaw

Files (1)

SKILL.md

2.5 KB

Realignment Meta-Framework

Shared framework for all realignment training workflows. Provides the common pipeline template, quality thresholds, and performance optimization guidance used across all domain-specific realignment workflows.

7-Stage Pipeline Template

All realignment workflows follow this common pipeline:

Capability Assessment: Evaluate current model capabilities and identify gaps
Data Preparation: Collect and prepare domain-specific training data
SFT Preparation: Supervised fine-tuning on curated examples
Preference/Reward Modeling: Domain-specific optimization (DPO, RLVR, SRF, etc.)
Iterative Training: Multi-round training with quality gates
Evaluation & Monitoring: Comprehensive evaluation against baselines
Deployment & Validation: Final validation and deployment readiness

Quality Thresholds

Metric	Minimum	Target	Critical
Task accuracy	85%	92%	< 80% triggers rollback
Capability retention	95%	98%	< 90% triggers rollback
Data quality score	0.8	0.9	< 0.7 blocks training
Evaluation coverage	80%	95%	< 70% blocks deployment

Capability Regression Detection

Run baseline evaluation suite before and after each training stage
Track per-capability scores across training rounds
Automatic rollback if any capability drops > 5% from baseline
Cross-domain contamination checks between training stages

Performance Optimization

Memory Management

Use gradient checkpointing for models > 7B parameters
Batch size auto-tuning based on available memory
Mixed precision training (fp16/bf16) by default

Training Efficiency

Learning rate warmup: 5-10% of total steps
Cosine annealing schedule with min_lr = 0.1 * max_lr
Early stopping with patience = 3 evaluation rounds
Checkpoint every N steps (configurable per domain)

Hardware Considerations

See mlx-performance skill for Apple Silicon optimization
GPU memory estimation: model_params * 4 bytes * 3 (model + optimizer + gradients)
Multi-device training coordination patterns

Cross-References

Hardware details: See mlx-performance skill
Domain workflows: See realign-domain-workflows skill
Data quality: See preference-data-quality skill

Source

git clone https://github.com/akaszubski/autonomous-dev/blob/master/plugins/autonomous-dev/skills/archived/realign-meta-framework/SKILL.md

View on GitHub

Overview

Provides a common 7-stage pipeline, quality thresholds, and optimization guidance for realignment workflows. It standardizes data prep, SFT, preference modeling, evaluation, and deployment across domains to ensure consistent quality and rollback safety.

How This Skill Works

The framework encodes realignment as a 7-stage pipeline: Capability Assessment, Data Preparation, SFT Preparation, Preference/Reward Modeling, Iterative Training, Evaluation & Monitoring, and Deployment & Validation. It defines concrete quality thresholds and automatic rollback rules, and bundles memory and training-efficiency guidance to optimize large-model workflows across domains.

When to Use It

Starting a new realignment domain workflow and defining the 7-stage pipeline
Enforcing quality gates and rollback policies across training stages
Training large models with memory and precision optimization (e.g., >7B params)
Detecting and preventing cross-domain contamination between training stages
Coordinating DPO/RLVR/SRF-style optimization within realignment cycles

Quick Start

Step 1: Define the realignment domain, data sources, and baseline capabilities
Step 2: Implement Data Preparation, SFT Preparation, and Preference Modeling with quality gates
Step 3: Run iterative training, monitor metrics, and perform deployment validation

Best Practices

Run baseline evaluations before and after each training stage to quantify impact
Define and document minimum, target, and critical metrics for each task
Enable automatic rollback if any capability drops more than 5% from baseline
Apply gradient checkpointing and mixed precision to improve memory and speed
Coordinate with related skills (mlx-performance, realign-domain-workflows, data-quality) for end-to-end quality

Example Use Cases

Realign a content moderation model using DPO with the 7-stage pipeline and rollback safeguards
Finance advisory assistant alignment applying data-quality thresholds and iterative gates
Medical QA model alignment ensuring capability retention above baseline with checks
Multimodal model alignment using gradient checkpointing on a large parameter set
Cross-domain realignment rollout across multiple teams with early stopping and gates

Frequently Asked Questions

Add this skill to your agents