Export for a consumer

Use an exact --profile to export for Hugging Face TRL (Transformer Reinforcement Learning), Fireworks, SWE-bench, or NeMo Gym. The Consumer compatibility reference lists the supported versions and checks. Without a profile, Sediment emits its canonical rows. Diff-SFT and Recovery remain Sediment-native. A profile changes representation at the consumer boundary. It preserves the Evidence recipe, labels, Reward, and Provenance. Missing negative examples remain missing. A permissive agent workflow can produce verified SFT rows while producing no DPO pairs. Prompt steering alone isn’t a rejection label.

Check consumer capacity

Profiles use the 64 MiB encoded materialization allowance from BundleLimits. SFT and DPO profiles check canonical training rows before collecting their consumer view. RLVR profiles check complete Rollouts before hydration; the NeMo check also includes the source Inference calls. A file-backed population that exceeds the allowance fails before its first record is decoded. The same allowance applies separately to projected native rows, adapted data and evidence together, loader inputs, consumer settings, and prepared data and evidence files together. Native projectors stage each admitted row before collecting a consumer view. Python callers that supply iterators receive a capacity error when their cumulative encoded rows exceed the allowance. These limits bound encoded populations, not total process memory. If a limit is exceeded, the complete export fails before publication. Sediment preserves the destination and removes its private staging. Capacity failures don’t become eligibility skips or smaller successful datasets. For a larger export, use canonical rows without --profile and qualify downstream conversion separately. The profiling procedure explains resource measurement and qualification boundaries.

Install the matching optional environment

Keep optional consumer packages separate from the API environment. For the Hugging Face profiles, create an isolated Python 3.12 environment and install the published Sediment release with the versions listed in the compatibility reference:

uv venv --python 3.12 "$HOME/.local/share/sediment-consumer"
uv pip install --python "$HOME/.local/share/sediment-consumer/bin/python" \
  'sediment-cli==0.3.0' 'datasets==5.0.1' 'trl==1.13.0' \
  'transformers==5.17.0' 'torch==2.14.0'
export PATH="$HOME/.local/share/sediment-consumer/bin:$PATH"

Fireworks format profiles need no optional package. They validate the format; they don’t upload data or run a hosted training job. The exact SWE-bench and NeMo profile qualifications use upstream revisions that aren’t supplied by the Sediment package. This guide doesn’t provide a package-only setup for those environments. Use canonical exports without --profile, or an independently qualified consumer environment that meets the reference contract.

Export SFT or DPO

If captured CI evidence supports verified imitation, select sft_verified:

sediment export sft --from /data/derived/review \
  --recipe sft_verified --profile hf-trl-sft-v1 --out /data/consumer/hf-sft

If captured human preferences support DPO, select dpo_human:

sediment export dpo --from /data/derived/review \
  --recipe dpo_human --profile hf-trl-dpo-v2 --out /data/consumer/hf-dpo

For Fireworks, use fireworks-sft-v1 or fireworks-dpo-v2. Every invocation requires an unused destination directory. An invalid profile, incompatible package version, invalid row, or incomplete configuration fails before publication. A populated export contains data, aligned evidence sidecars, and a compatibility manifest. An empty export reports zero rows and creates no directory. Preserve the sidecars for auditing; send only data to the consumer. DPO profiles v2 accept recipe-v2/schema-v4 rows. The retired DPO profile names v1 fail explicitly. If you retained recipe-v1/schema-v3 rows, re-export canonical evidence before selecting a successor; the adapter refuses historical rows. See Migrate retained DPO exports.

Load Hugging Face data

Use Sediment’s loader to preserve heterogeneous nested message and tool values. Inferring Arrow structs from early rows can reject a later tool argument shape.

from sediment_export.compatibility import load_hf_dataset

train = load_hf_dataset(
    "/data/consumer/hf-sft/data.train.jsonl", "hf-trl-sft-v1"
)

The loader uses explicit List(Json()) features and verifies an exact value round trip through datasets.Dataset. SFT retains separate prompt and completion fields. DPO retains prompt, chosen, and rejected. The adapter renames readable thinking to reasoning_content. It preserves structured tool arguments and emits structured results as deterministic JSON strings. Configure the trainer to respect the completion loss boundary. Confirm that your model’s chat template supports reasoning and tool calls before training. Qualification exercises TRL’s public conversation helper and the Qwen3-0.6B tokenizer at revision c1899de289a04d12100db370d81485cdf75e47ca. It doesn’t qualify every tokenizer or run an optimizer. See the upstream Dataset formats.

Interpret Fireworks output

Both Fireworks profiles reject developer messages and require any system message to appear first. Sediment doesn’t rewrite or reorder captured roles. The SFT adapter sets earlier assistant messages to weight: 0 and target assistant messages to weight: 1. These values mark the loss boundary. The adapter doesn’t emit a root sample weight or derive one from Confidence. Tool arguments become JSON strings. Readable reasoning uses reasoning_content. Tool definitions remain absent because the Fact doesn’t capture them. Configure any definitions required by your training model from an authoritative source. The DPO profile accepts the documented text-only message contract and one final assistant message per preferred and non-preferred output. Unsupported roles or structured DPO messages fail visibly. Fireworks requires at least three examples per uploaded dataset; a smaller local export isn’t evidence of hosted readiness. See the provider’s SFT format and DPO format.

Export NeMo Gym Rollouts

Supply a JSON configuration with explicit adapter parameters:

{
  "schema_version": 1,
  "nemo": {"parallel_tool_calls": false, "tool_choice": "auto", "tools": []}
}

These values describe your consumer setup. The evidence sidecar identifies the operator as their source; they aren’t reconstructed inference-request Facts.

sediment export rlvr --from /data/derived/review \
  --target nemo-gym --profile nemo-gym-rollouts-v1 \
  --consumer-config /data/nemo.json --out /data/consumer/nemo

The profile hydrates full captured messages through validated Inference-call references. It preserves Segment boundaries, ordered native Responses items, and tool-call links. It never substitutes scoring text for model output. NeMo Gym 0.2.1 stores readable reasoning in summary_text; the adapter copies the full text without summarizing it. The sidecar also retains native output messages. Response timestamps cite the source Inference call. Missing status and token usage stay absent. Every admitted row passes the upstream BaseVerifyResponse parser without silent field loss. The profile requires captured numeric Reward. It counts missing Reward, missing Inference calls, identity mismatches, missing or mixed models, and unsupported message mappings. The closed vocabulary lives in consumer_rlvr.py::NEMO_PROFILE_SKIP_REASONS. This is a native parser contract. A particular offline training recipe may require a further dataset formatter and tokenizer. In particular, the NeMo offline tutorial shows a formatting snippet with top-level output; this profile retains the native verification envelope’s response.output. Don’t pass that envelope to the snippet unchanged.

Supply SWE-bench task and runtime inputs

The SWE-bench profile requires an explicit repair request and executable task configuration for each canonical instance ID. First inspect the canonical --target swe-bench projection and its repository/base-commit binding. Then supply tasks in the consumer JSON configuration. Replace the illustrative values with your verified task configuration:

{
  "schema_version": 1,
  "tasks": {
    "exact-exported-instance-id": {
      "repo": "owner/repository",
      "base_commit": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa",
      "problem_statement": "The repair request you verified.",
      "test_patch": "",
      "hints_text": "",
      "created_at": "2026-09-01T00:00:00Z",
      "version": "your-task-version",
      "environment_setup_commit": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa",
      "FAIL_TO_PASS": ["test_name_that_fails_before_the_patch"],
      "PASS_TO_PASS": [],
      "image": "your-prepared-test-image",
      "eval_script": "echo '>>>>> Start Test Output'\npython -m pytest -rA\necho '>>>>> End Test Output'",
      "log_parser": "parse_log_pytest",
      "eval_type": "pass_and_fail"
    }
  }
}

Use an empty test_patch only when the required tests already exist at the base commit. FAIL_TO_PASS must name at least one test; its names can’t overlap PASS_TO_PASS. The runtime image must contain the checked-out repository and its dependencies. Match the command, output markers, and registered parser to the tests you run. Sediment validates the fields and parser registration but doesn’t run the image or establish test transitions.

sediment export rlvr --from /data/derived/review \
  --target swe-bench --profile swe-bench-tasks-v1 \
  --consumer-config /data/swe.json --out /data/consumer/swe

The repository and base commit must match the source row exactly. The patch remains the captured passing reference patch. The sidecar preserves the captured first user message separately from your explicit repair request and records the configuration digest. Agent instructions or environment context mustn’t become a guessed repair request. Qualification uses the pinned SWE-bench loader and runtime specification.

Check suitability before training

Inspect the manifest’s split counts, numeric Reward counts, and exclusions. skipped records consumer-adapter exclusions. CLI exports also preserve canonical_skipped from the training projection. RLVR exports preserve fragmented from the source bundle. These fields count separate populations; don’t add them together. Low-level publication omits a source population when the caller doesn’t supply it. A supplied empty map means zero recorded counts. An all-positive corpus can support imitation; parser acceptance doesn’t create useful preference pairs or Reward contrast. An empty labeled eval partition can’t measure held-out performance. Split exports reject identical full prompts shared by train and eval, but different prompts can describe the same task. Review task overlap and model context limits separately. The pinned CI matrix qualifies every advertised profile and gates release preparation. The weekly canary tests updated upstream packages without changing product pins. Dependency or Python-version incompatibility counts as drift. Fireworks has no local dataset loader; its format tests remain separate from hosted acceptance. If an upstream release needs different output bytes, add a profile version before changing the support claim.

Getting started

Agents

Concepts

Operate

Capture

Derive

Exports

Reference

Export for a consumer

Check consumer capacity

Install the matching optional environment

Export SFT or DPO

Load Hugging Face data

Interpret Fireworks output

Export NeMo Gym Rollouts

Supply SWE-bench task and runtime inputs

Check suitability before training

​Check consumer capacity

​Install the matching optional environment

​Export SFT or DPO

​Load Hugging Face data

​Interpret Fireworks output

​Export NeMo Gym Rollouts

​Supply SWE-bench task and runtime inputs

​Check suitability before training

Check consumer capacity

Install the matching optional environment

Export SFT or DPO

Load Hugging Face data

Interpret Fireworks output

Export NeMo Gym Rollouts

Supply SWE-bench task and runtime inputs

Check suitability before training