> ## 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.

# Export SFT and diff-SFT rows

> Export eligible responses or exact attributed-file patches.

For a supported downstream release, select a versioned
[consumer profile](/exports/consumer-compatibility).

Run the default supervised fine-tuning (SFT) and diff-shaped SFT (diff-SFT)
exports:

```bash theme={"dark"}
sediment export sft --out /data/export
sediment export diff-sft --out /data/export
```

The commands write `sft.jsonl` and `diff_sft.jsonl`, or split train and eval
files. Both select `sft_curated` version 1 by default. For shared prerequisites,
reviewed-bundle steps, output inspection, and holdout guidance, see [Choose a
training export](/exports/training-exports).

## Choose an Evidence recipe

SFT and diff-SFT use the same selected Evidence recipe and `SFTPolicy`.
Diff-SFT doesn't redefine eligibility.

`sft_curated` version 1 is the default. It admits a human-explicit accept or an
accepted edit with `edit_retention_score >= 0.8`. A human-explicit reject,
abandonment, or resolved workflow failure vetoes eligibility. The veto includes
an ambiguous aggregate with one failing workflow. A resolved CI pass can affect
Confidence and CI reliability, but it can't create curated eligibility.

To admit only completions with a clean resolved CI pass, select the verified
recipe:

```bash theme={"dark"}
sediment export sft --recipe sft_verified --out /data/export
sediment export diff-sft --recipe sft_verified --out /data/export
```

`sft_verified` version 1 requires explicit selection. It admits only a clean
resolved CI pass. Ambiguous verdicts, non-verdict-only evidence, and suspected
flakes don't create eligibility.

Both recipes require a resolved Confidence of at least
`SFTPolicy.min_confidence`, which defaults to 0.6.

## Export SFT rows

The SFT export writes at most one row per inference call that the selected
recipe says is safe to imitate.

An inference call can have one eligible Attributed completion per attributed
file. The projection keeps the one with the highest Confidence, then uses the
repository-qualified evidence identity to break ties. It counts identical extra
copies as `duplicate_completion`. If copies of one complete evidence identity
carry conflicting payloads, the projection declines the entire Inference call
once under `conflicting_evidence`. A higher-ranked sibling cannot hide the
conflict. This includes conflicts in split, Provenance, and source Fact content.

### Read an SFT row

| Field        | Meaning                                                                                                                                       |
| ------------ | --------------------------------------------------------------------------------------------------------------------------------------------- |
| `prompt`     | trainer-facing request messages; earlier assistant responses remain here, outside the loss boundary                                           |
| `completion` | the latest response-message list and the only SFT target                                                                                      |
| `tools`      | tool definitions; empty because inference-call version 1 doesn't carry definitions                                                            |
| `metadata`   | organization, source model, completion ID, recipe ID and version, eligibility source, label Confidence, CI reliability, Provenance, and split |

Example SFT row:

```json theme={"dark"}
{
  "prompt": [
    {"role": "user", "content": "Inspect the file."},
    {"role": "assistant", "tool_calls": [{"id": "tool-1", "type": "function", "function": {"name": "Read", "arguments": {"path": "app.py"}}}]},
    {"role": "tool", "name": "Read", "tool_call_id": "tool-1", "content": "def run(): ..."},
    {"role": "user", "content": "Fix it."}
  ],
  "completion": [{"role": "assistant", "thinking": "Keep the signature.", "content": "Implemented."}],
  "tools": [],
  "metadata": {
    "org_id": "acme", "source_model": "model-name", "completion_id": "call-1",
    "recipe_id": "sft_curated", "recipe_version": 1,
    "eligibility_source": "explicit_accept",
    "label_confidence": 0.9,
    "ci_reliability": 1.0,
    "provenance": {"policy_version": "5", "quarantine_revision": 0, "policy_digest": null},
    "attribution_source": "jaccard",
    "session_commit_observation_ids": [],
    "split": "train",
    "repository_identity": null,
    "schema_id": "https://sediment.so/schemas/training-rows/sft-sample/v3.json",
    "schema_version": 3
  }
}
```

The mapper keeps text in `content`, readable reasoning in `thinking`, and tool
calls in `tool_calls`. Arguments remain objects. Tool results retain their order;
structured results become deterministic JSON strings. Earlier assistant messages
stay in `prompt`, outside the completion loss boundary.

### Interpret skipped SFT inputs

Inspect `skipped` before training. The SFT export reports every skipped input
under this closed vocabulary: `conflicting_run_identity`,
`ambiguous_workflow_verdicts`, `repository_identity_absent`,
`repository_identity_conflict`, `repository_identity_unresolved`,
`repository_mirror_identity_unresolved`, `repository_source_absent`,
`non_finite_number`, `unrepresentable_unicode`, `completionless`,
`duplicate_tool_call_id`, `empty_message`, `non_string_tool_result`,
`unrepresentable_part_order`, `unresolved_tool_call`,
`unsupported_completion_role`, `unsupported_message_role`,
`unsupported_role_part`, `abandoned`, `explicit_reject`, `resolved_ci_failure`,
`no_eligibility_source`, `unreliable_ci_resolution`, `no_reward_signal`,
`below_confidence_floor`, `inference_call_not_found`, `model_absent`,
`duplicate_completion`, and `conflicting_evidence`.

## Export diff-SFT rows

The diff-SFT projection writes one row per `(inference call, commit)` group and
reads the concrete per-file edits from the mirror. It never synthesizes a diff.

The projection skips abandonment rows before grouping or mirror access. If any
remaining member is ineligible or explicitly rejected, the projection skips
the whole group. Group Confidence is the `min()` across members.

### Read a diff-SFT row

| Field        | Meaning                                                                                                                                                                               |
| ------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `prompt`     | the trainer-facing request messages                                                                                                                                                   |
| `completion` | one assistant message whose `content` is the exact selected unified patch                                                                                                             |
| `tools`      | tool definitions; empty because inference-call version 1 doesn't carry definitions                                                                                                    |
| `metadata`   | organization, Session, repo, commit, source model, completion ID, recipe ID and version, eligibility source, label Confidence, CI reliability, Provenance, split, and source Fact IDs |

The `completion` value uses this shape, with the mirror text preserved exactly:

```json theme={"dark"}
[{"role": "assistant", "content": "diff --git a/app.py b/app.py\n--- a/app.py\n+++ b/app.py\n@@ -1 +1 @@\n-old\n+new\n"}]
```

The projection selects each attributed file's complete patch section from the
mirror bytes and concatenates sections in the commit diff's order. If any
attributed file has no patch section, the projection skips the whole sample.
It doesn't synthesize patch headers or emit `files[].added_lines`. The shared
Git parser decodes quoted filenames while preserving the original patch text,
including rename, deletion, empty-file, and no-newline-marker sections. Invalid
sections never contribute partial patches. A declined section counts once per
commit even when several groups use that commit; valid selected sections can
still produce rows.

### Interpret skipped diff-SFT inputs

The diff-SFT export reports every skipped input under this closed vocabulary:
`conflicting_run_identity`, `ambiguous_workflow_verdicts`,
`repository_identity_absent`, `repository_identity_conflict`,
`repository_identity_unresolved`, `repository_mirror_identity_unresolved`,
`repository_source_absent`, `non_finite_number`, `unrepresentable_unicode`,
`completionless`, `duplicate_tool_call_id`, `empty_message`,
`non_string_tool_result`, `unrepresentable_part_order`, `unresolved_tool_call`,
`unsupported_completion_role`, `unsupported_message_role`,
`unsupported_role_part`, `abandoned`, `explicit_reject`, `resolved_ci_failure`,
`no_eligibility_source`, `unreliable_ci_resolution`, `no_reward_signal`,
`below_confidence_floor`, `inference_call_not_found`, `model_absent`,
`duplicate_completion`, `conflicting_evidence`, `unsupported_diff_section`,
`malformed_diff_section`, `repo_mismatch`, `split_mismatch`,
`eligibility_source_mismatch`, `mirror_absent`, `commit_diff_unavailable`,
`empty_patch`, and `file_diff_unavailable`. `unsupported_diff_section` and
`malformed_diff_section` identify unusable Git evidence.

Keep Attribution source metadata and observation IDs with each row. Version 1
recipe eligibility permits inferred Attribution; an empty observation list means
that the row carries no observed Session-to-commit evidence. See
[the source contract](https://github.com/sediment-ai/sediment/blob/main/docs/adr/0014-factual-outcomes-and-training-evidence.md#evidence-recipes-and-exact-metadata).