Skip to main content
Export captured Rollouts as task rows, trajectory rows, or both for reinforcement learning from verifiable rewards (RLVR). This procedure is for operators who need to audit trajectories or prepare a downstream dataset. The exporter doesn’t run a Verifier or fill missing evidence.
  1. Choose a target.
  2. Prepare the inputs.
  3. Optional: Configure verification.
  4. Export one target.
  5. Inspect the output.

​
Choose a target

--target is required. These are Sediment’s canonical projections; a target name alone doesn’t establish downstream compatibility. If you need a qualified SWE-bench or NeMo Gym release, use Export for a consumer after choosing the target. Every emitted row names version 1 of the rlvr_ci Evidence recipe. A resolved CI pass records reward_source: resolved_ci_pass. A resolved CI failure records reward_source: resolved_ci_fail. A non-verdict omits reward_source and doesn’t create a directional label or fractional Reward.

​
Prepare the inputs

Build and review a bundle, then prepare a private output parent. The examples use $HOME/sediment-derived/review. For sediment or swe-bench, set SEDIMENT_MIRROR_PATH to the deployment’s mirror root. The mirrors must contain the attributed commits, their parents, and the exact verified commit. The bundle freezes Rollouts, not Git patches. For nemo-gym with --from, no mirror or database connection is required. If you omit --from, configure direct Fact access. Every direct RLVR target needs database and mirror access.

​
Configure verification

Verification configuration tells a consumer how a Verifier can run again. A Verifier result is the exact recorded CIOutcome Fact. Every target keeps these values separate. If a repository has a known verification command, save it in $HOME/sediment-exports/verifier-commands.toml. Replace owner/repo and the example command with that repository’s verified values:
Set the configuration path:
If the setting or repository entry is absent, the projection omits verification. The projection retains recorded Verifier results. Sediment never derives a command from workflow YAML, and it never runs the configured command. If you need the Sediment target’s experimental environment.yaml, set its operator-supplied runtime inputs before exporting. That manifest doesn’t establish an executable environment.

​
Export one target

Run the command for the target you selected. Each example uses a different, unused destination:
For later runs, choose another unused directory. Don’t rebuild the reviewed bundle as part of the export step.

​
Inspect the output

The command prints rollouts derived, task rows, rollout rows, separate skip maps, fragmented, and the written paths.
  1. Check the selected target’s row counts. A nonempty Rollout population can produce no tasks when Verifier or patch evidence is missing.
  2. Inspect recorded Verifier results separately from their resolved verdict, Reward, and reliability. A numeric Reward alone doesn’t establish reliable evidence.
  3. Review fragmented before treating separate Segments as one trajectory. The counts describe retained continuity boundaries, not excluded rows.
  4. Review target-specific skips and audit the split.
With the default evaluation split, the writer produces <name>.train.jsonl and <name>.eval.jsonl. Otherwise, it produces one <name>.jsonl. An empty row list doesn’t create or truncate a file. The built-in split is Session-grained. It keeps every segment and task derived from one Session on one side. It doesn’t isolate repositories, exact prompts, or external task identifiers across Sessions. If a benchmark uses task-keyed isolation, audit both files against the versioned task manifest before training or evaluation. Use a fresh output directory for each export to keep task, Rollout, split, and manifest files from different runs separate. The first export writes a hidden .sediment-rlvr-target claim in its output directory. If the directory contains an RLVR artifact, another export fails before writing, including an export for the same target. A directory with unclaimed RLVR artifacts also fails closed. An empty export can reuse its claimed directory because it wrote no artifact. A hidden generation lock rejects concurrent exports for the same directory. If a stopped process leaves the lock behind, the directory remains fail-closed. Use a fresh directory for the next export.

​
Understand target contracts

Compare repeated exports within one mode by JSONL bytes and row order. Across direct and bundle exports, compare parsed values and order: bundles sort nested keys while direct exports retain source order. Inspect representation skips such as non_finite_number and unrepresentable_unicode. Each declined task or Segment counts once. File replacement is atomic per file, not across all partitions. If publication fails or stops, retain its diagnostics and export to a fresh directory. Empty partitions don’t clear existing files. Validate rows against the Schema reference before consumer adaptation. A target contract doesn’t establish compatibility with an external trainer release; use a qualified consumer profile. Canonical schema, Evidence recipe, and consumer-profile versions identify separate contracts.

​
Read Sediment target rows

The target writes task rows and Rollout rows. Task rows preserve a passing or failing historical patch, recorded Verifier results, CI resolution, Attribution source, observation IDs, split, and Provenance. The Reference patch starts at the parent of the first attributed commit in the selected CI resolution’s repository and ends at its exact verified commit. Later unverified commits and commits in other repositories don’t extend it.

​
Resolve attempts

Inspect ci_resolution alongside the exact verifier_results. CI evidence and verdicts explains attempt ordering, workflow conflicts, and suspected flakes. A pass after a failure, or a failure after a pass, can carry a categorical verdict with default reliability 0.0. NeMo Gym Reward remains 1.0, 0.0, or absent; reliability doesn’t change that numeric value. If you weight or filter training rows by reliability, define that downstream rule explicitly.

​
Read task prompts

Task exporters retain each canonical TextPart.content verbatim, including JSON-looking examples, whitespace, and Unicode. They join explicit parts with newlines and mark each non-text part as [non-text content omitted]. They don’t decode text as legacy provider content. In canonical rollouts.jsonl, each Turn preserves typed input messages in new_messages. Its completion contains scoring text: joined output text and tool-argument string values. That field omits readable reasoning and output part boundaries. Full captured responses remain in the bundle’s Inference calls. If you need those responses in NeMo’s native format, use the NeMo consumer profile.

​
Read Rollout rows

Rollout implementation version 4 continues a Segment only when typed input replays the prior input and captured output. Missing or contradictory echoes, rewritten input, missing output, and unsupported message regrouping retain the Turn in a separate Segment. The Derivation counts these boundaries in fragmented, separately from projection skips. Each retained Rollout Segment becomes one row. The row preserves Session identifiers, segment index, structured messages, completion text, tool calls, decisions, split, Provenance, and exact verifier_results. Session-terminal Verifier results repeat on each segment because Sediment has no Fact that assigns a result to one segment. ci_resolutions carries the recomputed commit-level interpretations. ci_resolution identifies the resolution that supplied reward_source, including its reliability. Check these task exclusions when diagnosing missing output: Sediment Rollout rows add no_segments to the shared CI-resolution, repository-identity, and representation exclusions.

​
Read SWE-bench target rows

See SWEBenchTaskRow for the complete shape. The row carries repository, base commit, problem statement, patch, and Sediment evidence in metadata. The target populates patch only from a recorded terminal pass. It omits issue identifiers, test patches, versions, FAIL_TO_PASS, and PASS_TO_PASS because Sediment Facts don’t contain them. SWE-bench adds failed_reference_patch to the Sediment task skip vocabulary. It counts that reason when the recorded terminal Verifier result is a failure. This count remains distinct from missing_verifier_evidence.

​
Read NeMo Gym target rows

Each captured Segment maps to a NemoGymRolloutRow with responses_create_params, response, optional reward, and evidence in metadata. A resolved pass maps to numeric Reward 1.0. A resolved failure maps to 0.0. Without a resolved pass or failure, the row omits both reward and reward_source. Metadata retains every exact terminal Verifier result, including non-label outcomes. It keeps the selected ci_resolution separate from optional Verification configuration. The target doesn’t infer a NeMo Gym resource server or executable environment. Its skips include no_segments and the shared CI-resolution, repository-identity, and representation exclusions. This mapping names NeMo Gym’s rollout boundary, but it doesn’t claim direct compatibility. The canonical target retains Turn input messages, scoring text, and tool calls. The NeMo consumer profile loads full output messages from the referenced Inference calls and validates the upstream response format.

​
Interpret the experimental environment manifest

Only the Sediment target can write environment.yaml. The document contains experimental: true because its OpenEnv shape follows a proposal rather than a compatibility-tested contract. SWE-bench and NeMo Gym exports never write this file. The manifest can include an informational CI workflow reference when every task agrees on one workflow. It includes frameworks.openenv only from SEDIMENT_OPENENV_IMAGE or SEDIMENT_OPENENV_PACKAGE. It includes frameworks.nemo_gym only from an operator-supplied SEDIMENT_NEMO_GYM_RESOURCES_SERVER, with optional SEDIMENT_NEMO_GYM_CONFIG. It never derives either runtime from workflow YAML. The OpenEnv field mapping documents intent, not compatibility. An OpenEnv dataset needs a runtime that implements the expected protocol. The Verifier-result-to-Reward mapping doesn’t provide a Verifiers taskset or Verifier implementation.

​
Understand capture limits

Rollout rows require captured Inference calls. The supported capture path for those Facts is the gateway. OpenTelemetry-only capture can produce decisions and CI Facts without a trajectory. Copilot model calls can’t route through the Sediment gateway, so Copilot contributes decisions but not Rollouts. For field-level types, see the generated Schema reference. For Python use, start with the validated bundle reader and sediment_export.export_rlvr_from_bundle. Keep the reader context open through projection.