Skip to content
Arxo Arxo

Fine-tuning Architecture

Fine-tuning Architecture

Section titled “Fine-tuning Architecture”

This guide shows an end-to-end workflow for improving finetuning_architecture health in production fine-tuning systems.

Workflow

Section titled “Workflow”
  1. Run a fine-tuning architecture audit.
  2. Triage findings by risk family and blast radius.
  3. Apply targeted remediation.
  4. Enforce policy gates in CI.
  5. Prevent regressions with baseline checks.

1) Run an Audit

Section titled “1) Run an Audit”
Terminal window
# Focused fine-tuning metric
arxo analyze --path . --metric finetuning_architecture --format json
Terminal window
# AI preset (includes finetuning_architecture)
arxo analyze --path . --preset ai --format json

2) Triage Findings

Section titled “2) Triage Findings”

Prioritize in this order:

  1. method_integrity_score, checkpoint_eval_lineage_score, artifact_trust_surface_score
  2. base_model_versioning_score, run_lineage_score, determinism_envelope_score
  3. eval_absence_score, dataset_contamination_score, distillation_integrity_score
  4. privacy_recordkeeping_score, model_artifact_access_score, adapter_isolation_score
  5. oom_risk_score, cost_tracking_score, artifact_metadata_score, prompt_format_inconsistency_score

Then inspect composite movement:

  • finetuning_architecture.overall_finetuning_health
  • finetuning_architecture.reproducibility_score
  • finetuning_architecture.data_integrity_score
  • finetuning_architecture.safety_governance_score

3) Apply Fixes by Track

Section titled “3) Apply Fixes by Track”
  • Method track: enforce profile-specific invariants (dpo, ppo, rft, grpo, rloo, distill).
  • Repro track: base/tokenizer pinning, run passport, determinism envelope, checkpoint-eval linkage.
  • Data/eval track: explicit split strategy, contamination checks, eval maturity, distillation provenance/parity.
  • Safety/governance track: trust surface hardening, access controls, privacy budget logging, technical recordkeeping.
  • Ops/cost track: OOM controls, budget and throughput tracking.

Use the Remediation Playbook for detector-by-detector fixes.

4) Enforce in CI

Section titled “4) Enforce in CI”

Use profiles from Policy and CI Gates.

Terminal window
arxo analyze --path . --preset ai --config arxo.yml --fail-fast

Recommended rollout:

  1. Start with warning-level thresholds.
  2. Fix recurring low-score categories in central modules.
  3. Promote critical gates to error after score trend stabilizes.

5) Baseline and Monorepo Rollout

Section titled “5) Baseline and Monorepo Rollout”
  • Enable baseline no-regression checks against origin/main.
  • Start in one critical training service/workspace.
  • Expand to additional workspaces after stable score trends.
Section titled “Related Reading”