Skip to content
posttrainllm docs
Esc
↑↓navigate↵open⌘Jpreview
On this page

SQL Technique Backlog

SQL Technique Backlog

This is the SQL-specific ledger of methods, recipes, status, and next smallest tests. It exists to prevent the roadmap from looking complete just because broad method names are present.

Current Evidence

Recipe / Attempt Status Evidence Decision
Toy SQL SFT, rank 4 Worked as a factory proof execution 0.167 -> 0.833 on 6 rows retry-data because train/eval overlap
Expanded synthetic SQL SFT, rank 4 Worked execution 0.160 -> 0.860, exact 0.140 -> 0.840 on 50 rows current synthetic specialist
Public b-mc2 SFT v1 Failed exact 0.344 vs T5-small 0.484 reject
Public b-mc2 SFT v2 Failed badly exact 0.031 reject
Public b-mc2 SFT v3 Improved but failed gate exact 0.422 reject
Public b-mc2 SFT v4, join/group weighted Worked on public exact exact 0.531 vs T5-small 0.484 not shippable; synthetic regression
Public v4 synthetic regression Failed synthetic execution 0.240 vs incumbent 0.860 route or compose
Public+synthetic blended SFT Failed/interfered public 0.297, synthetic execution 0.560 reject
Static multi-LoRA composition Failed to pass both gates best public pass still had synthetic failure reject
Routed public + synthetic adapters Best current candidate public exact 0.531, synthetic execution 0.860 report-ready, not package-ready
Hygiene SimPO/DPO adapter Failed hard execution 0.860 -> 0.080, degenerate output retry-training

Not Yet Tried

Technique Method Concrete SQL recipe Why It Might Help Smallest Next Test
Candidate-selection curriculum supervised selection / RLVR bridge Build rows with 4-6 candidate SQL strings, train/eval the model to pick the executable/gold-equivalent answer before open generation SQL generation is sparse-reward; selecting among candidates is easier and creates denser learning signal Use scripts/build_sql_candidate_choice.py on existing candidate predictions, create a tiny SFT/eval slice, compare selection accuracy by join/filter/group slices
Reference-anchored hygiene DPO DPO Retry the failed hygiene goal with a reference anchor, lower LR, fewer steps, and composed eval against the SFT adapter Ref-free SimPO over-optimized and collapsed; reference anchoring should preserve generation quality Same 108 pairs, frozen 50-row eval, require no execution regression and clean-SQL lift
One-step offline rollout update OAPL/ReST-style batch loop Generate N rollouts per prompt, score offline by execution/gold/format, train one adapter update, evaluate heldout Keeps the loop batch-first and avoids moving-target on-policy instability Render a batch plan with scripts/render_batch_posttrain_plan.py, then run only after candidate-selection evidence
Policy lag / stale reference RL regularization Keep rollout policy fixed for a batch or compare against a stale reference during preference/RL update Prevents the optimizer from chasing its own malformed outputs Add as a variant only after a first offline rollout update exists
Controlled LoRA rank sweep LoRA/DoRA Sweep rank {1,2,4,8} on the same frozen SQL data with fixed seed/steps/LR Existing ranks were confounded with different data/recipes; successful adapters have low effective stable rank Run no more than one small sweep after the next target is frozen; report slice metrics and geometry
LoRA geometry decision check diagnostics Compare successful, failed, and retry adapters by effective-update norm/stable rank/module concentration Explains whether a failure learned too little, too diffusely, or in the wrong layers Run scripts/lora_geometry.py on every meaningful adapter and attach lora-geometry.json
Slice-gated reporting eval discipline Require overall, join, single-table, filter, aggregate, format, and clean-output slices Overall hides the known join weakness and hygiene failure Generate slice-metrics.json for every SQL report
Trace review as data source failure analysis Classify failures into hallucinated schema, missing join, wrong filter, prose/fence wrapping, no-select collapse Converts failed attempts into targeted data or preferences Generate trace_review.md for every SQL report

Priority Order

  1. Candidate-selection curriculum. This is the cheapest new recipe and directly addresses sparse SQL reward.
  2. Reference-anchored hygiene DPO. This retries the known hygiene failure with a safer recipe.
  3. Public execution gate. This changes the eval from exact string match to a more serious SQL claim.
  4. Controlled rank sweep + geometry. Run after the next target is frozen so the sweep is not confounded.
  5. Offline rollout update / OAPL-style loop. Run only after candidate-selection creates a clean reward surface.

Enforcement

No SQL candidate should be reported as improved unless the run folder contains:

  • eval-baseline.json
  • eval-candidate.json
  • slice-metrics.json
  • trace_review.md
  • report.md
  • decision.json

No SQL candidate should be packaged unless the decision is ship and the public execution gate is present or explicitly waived in the report.

Was this page helpful?