> ## Documentation Index
> Fetch the complete documentation index at: https://docs.sediment.so/llms.txt
> Use this file to discover all available pages before exploring further.

# Consumer compatibility reference

Generated by `scripts/gen_compatibility_docs.py` from
`packages/export/sediment_export/compatibility.py`. Don't edit this page.

A canonical schema validates Sediment's contract. A consumer profile
validates the exact boundary named here. Neither proves model quality,
task correctness, or hosted job acceptance. Profile versions don't replace
Evidence recipe versions, canonical schema versions, or Provenance.

| Profile                | Objective | Version | Canonical input                                                      | Supported upstream                                       | Support                                             |
| ---------------------- | --------- | ------- | -------------------------------------------------------------------- | -------------------------------------------------------- | --------------------------------------------------- |
| `hf-trl-sft-v1`        | SFT       | 1       | `https://sediment.so/schemas/training-rows/sft-sample/v3.json`       | `datasets==5.0.1`, `trl==1.13.0`, `transformers==5.17.0` | loader                                              |
| `hf-trl-dpo-v2`        | DPO       | 2       | `https://sediment.so/schemas/training-rows/dpo-pair/v4.json`         | `datasets==5.0.1`, `trl==1.13.0`, `transformers==5.17.0` | loader                                              |
| `fireworks-sft-v1`     | SFT       | 1       | `https://sediment.so/schemas/training-rows/sft-sample/v3.json`       | Fireworks documentation, 2026-09-13                      | format; hosted acceptance not tested                |
| `fireworks-dpo-v2`     | DPO       | 2       | `https://sediment.so/schemas/training-rows/dpo-pair/v4.json`         | Fireworks documentation, 2026-09-13                      | format; hosted acceptance not tested                |
| `swe-bench-tasks-v1`   | RLVR      | 1       | `https://sediment.so/schemas/training-rows/swe-bench-task/v3.json`   | `swebench==5.0.1`, `datasets==5.0.1`                     | loader and runtime parser; repository tests not run |
| `nemo-gym-rollouts-v1` | RLVR      | 1       | `https://sediment.so/schemas/training-rows/nemo-gym-rollout/v4.json` | `nemo-gym==0.2.1`, `openai==2.6.1`                       | native parser; optimizer not tested                 |

## Validation entry points

| Profile                | Validation                                                     |
| ---------------------- | -------------------------------------------------------------- |
| `hf-trl-sft-v1`        | `datasets.Dataset.from_list; trl.is_conversational`            |
| `hf-trl-dpo-v2`        | `datasets.Dataset.from_list; trl.is_conversational`            |
| `fireworks-sft-v1`     | `Fireworks documented JSONL contract, 2026-09-13`              |
| `fireworks-dpo-v2`     | `Fireworks documented JSONL contract, 2026-09-13`              |
| `swe-bench-tasks-v1`   | `swebench.harness.utils.load_swebench_dataset; make_test_spec` |
| `nemo-gym-rollouts-v1` | `nemo_gym.base_resources_server.BaseVerifyResponse`            |

## Output artifacts

Every populated profile publishes a private directory containing
`data.jsonl`, `evidence.jsonl`, and `compatibility.json`. If splitting
is enabled, populated partitions use `data.train.jsonl` and
`data.eval.jsonl`, with matching evidence files. Empty partitions
produce no file. An empty export creates no destination.

The manifest records file hashes, profile identity, skip counts, split
counts, numeric Reward counts, and exact prompt overlap. The evidence
sidecar preserves canonical metadata and source hashes. It isn't trainer
input. Publication refuses an existing destination and validates every
row before exposing the directory. Split exports reject exact prompts
shared by train and eval; semantic task overlap still needs review.

`skipped` counts adapter exclusions. CLI profiles preserve
`canonical_skipped` from their training projection; RLVR profiles also
preserve source-bundle `fragmented` counts. These populations stay
separate. Low-level callers omit unavailable source populations; an
explicitly supplied empty map records zero counts.

See [Export for a consumer](/exports/consumer-compatibility) for
installation, exact commands, configuration, and consumer limitations.