Skip to main content
Choose an objective, prepare its evidence, and audit the output before training. For a supported trainer release, use a versioned consumer profile. Sediment exports direct preference optimization (DPO), supervised fine-tuning (SFT), diff-shaped SFT (diff-SFT), Recovery, and reinforcement learning from verifiable rewards (RLVR) rows.

​
Choose an objective

Rows retain qualified repository identity; DPO keeps each member’s identity separately. Renames preserve identity. Matching names or commits don’t join different repository lifetimes. Inspect repository skip counts for absent or conflicting evidence.

​
Check representation exclusions

Inspect non_finite_number and unrepresentable_unicode in export skip counts. Every objective excludes a row if an emitted value contains a non-finite number or an unpaired surrogate, including nested tool arguments, keys, and metadata. A DPO or Recovery pair counts once when either side fails. Source content that the Evidence recipe omits doesn’t exclude the row. Facts and canonical bundles preserve those values; exports don’t replace them with null or replacement text. See Training representation.

​
Prepare the export

Set SEDIMENT_ORG_ID. Direct exports need database and mirror access. Bundle projections have these requirements: The canonical Derivation policy defaults eval_fraction to 0.1. Run the installed CLI in the configured operator shell. Use private output and staging directories with enough space. For example:

​
Export a reviewed bundle

For a reviewed, reproducible workflow, derive once. Then project the same bundle into each per-completion format:
For RLVR, pass the same bundle and an explicit target:
The sediment target still reads mirrors for Reference patches. Recovery reads CI Facts and mirrors directly and doesn’t accept --from. See Run derivations for policy, cohort selection, inspection, and recomputation.

​
Inspect the output

The DPO, SFT, diff-SFT, and Recovery exports write dpo.jsonl, sft.jsonl, diff_sft.jsonl, or recovery.jsonl. If eval_fraction is greater than zero, the exporter writes <name>.train.jsonl and <name>.eval.jsonl instead. Writes are atomic. If a projection is empty, the exporter leaves existing files untouched and prints nothing written. It never truncates an earlier export. Every run prints row counts and a skipped tally. Each format uses a closed skip-reason vocabulary, so no ineligible input disappears silently. The RLVR writer has additional output-directory safeguards. Inspect the RLVR output before you reuse an RLVR export directory. DPO, SFT, and diff-SFT rows keep Sediment evidence under metadata, outside the trainer inputs. For those rows, metadata.split contains train or eval. Recovery uses its Sediment-native row envelope. RLVR uses the location that its target contract defines. Structured provenance contains policy_version, integer quarantine_revision, and nullable full policy_digest. Confidence values come from the Confidence ladder. The eval split is deterministic per Session (Eval split). Validate each row’s canonical schema_id and schema_version before adapting it. The Schema reference defines fields; the consumer profile versions the downstream mapping. Schema, recipe, policy, and database versions identify separate contracts. Confidence and CI reliability don’t set trainer weights. If you use them for weighting, define and version that mapping with the Evidence recipe and source. A Confidence of 0.9 doesn’t mean that 90% of such rows are correct. Before you treat Confidence as a probability, check it against independent judgments.

​
Audit the split before training

If you plan to train on an export, run dataset diagnostics with the same Evidence recipes first:
The report counts model and recipe balance, Confidence distributions, cross-split exact prompt duplicates, near-duplicate DPO prompt buckets, SFT Confidence-floor exclusions, sparse DPO buckets, abandonment coverage, and diagnostic final Fate counts. Fate doesn’t enter an Evidence recipe or any training row. If you capture pull request merge Facts, run sediment report merge-retention to inspect whether attributed changes survived review and integration. Its thresholds are sensitivity diagnostics. They don’t enter an Evidence recipe, Confidence, Reward, or any training row. If you need to audit individual scored changes, write the canonical rows while you generate the aggregate report:
The JSONL artifact contains source identities, boundary scores, Attribution metadata, and Provenance. It contains no source text, prompt text, completion text, Developer decision label, training eligibility, Reward, Confidence, or Evidence recipe. An empty or failed export leaves an earlier destination file untouched. The built-in split keeps one Session in one partition. It doesn’t enforce a repository, prompt, or task-keyed holdout across Sessions. If a benchmark uses one of those boundaries, compare every exported row with the benchmark manifest through its source identifiers. Exclude pilot or development tasks before training. Preserve the diagnostics, benchmark manifest, Derivation manifest, and export files together. The row-level split value alone doesn’t prove that the evaluation set matches a stronger holdout claim.

​
Retain source metadata

Keep each row’s Attribution source and Session observation ID fields when you prepare trainer inputs. Recipes permit inferred Attribution; an empty observation list leaves observed identity unavailable. Full matching Facts live on the canonical artifacts. ADR 0014 defines the source scope for both DPO members, SFT and diff-SFT targets, Recovery enrichment sides, and each RLVR target.