Datasets:

Heisenger
/

sft-benchmark-trajectories

# Source audit method

The current benchmark-by-benchmark result is
[`trajectory_sources.md`](trajectory_sources.md); machine-readable source and
measured-library coverage lives in `data/registry/sources.json` and
`data/coverage/trajectory_sources.json`.

Each hosted release was checked at an immutable commit SHA. Parquet imports
first verify that the repository API still reports that SHA, then download
revision-qualified source files, hash every shard, stream every row and retain
row-level rejects. Native JSON/JSONL releases use file hashes plus the same
adapter/validation gate. Dataset-server previews were used only for discovery,
never as evidence of a complete large cell.

The audit keeps four findings distinct:

1. a full trajectory release was found and imported;
2. a qualifying release exists but was not or could not be fully imported;
3. the inspected artifact exists but is not a trajectory (for example, final
   calls, gold answers or reference scripts);
4. no qualifying release was found after searching, which is not a claim of
   nonexistence.

An imported record must preserve its raw source, revision, record locator,
ordered decisions and source-derived observations. At least two decisions with
complete intervening context are required for structural SFT eligibility.
Partial traces may remain as explicitly tier-5 diagnostics. Final-answer-only
records never enter the canonical trajectory population.

Outcome and provenance are independent. Human/reference demonstrations,
recorded model rollouts, source evaluator verdicts and local reproduction are
never collapsed into one “expert” flag. No trajectory is called optimal.

Licenses and training restrictions are recorded even though this task prepares
data rather than training a model. Evaluation-only sources are retained for
baseline analysis and assigned `protected_evaluation`; they are ineligible for
default SFT export.

	# Source audit method

	The current benchmark-by-benchmark result is
	[`trajectory_sources.md`](trajectory_sources.md); machine-readable source and
	measured-library coverage lives in `data/registry/sources.json` and
	`data/coverage/trajectory_sources.json`.

	Each hosted release was checked at an immutable commit SHA. Parquet imports
	first verify that the repository API still reports that SHA, then download
	revision-qualified source files, hash every shard, stream every row and retain
	row-level rejects. Native JSON/JSONL releases use file hashes plus the same
	adapter/validation gate. Dataset-server previews were used only for discovery,
	never as evidence of a complete large cell.

	The audit keeps four findings distinct:

	1. a full trajectory release was found and imported;
	2. a qualifying release exists but was not or could not be fully imported;
	3. the inspected artifact exists but is not a trajectory (for example, final
	calls, gold answers or reference scripts);
	4. no qualifying release was found after searching, which is not a claim of
	nonexistence.

	An imported record must preserve its raw source, revision, record locator,
	ordered decisions and source-derived observations. At least two decisions with
	complete intervening context are required for structural SFT eligibility.
	Partial traces may remain as explicitly tier-5 diagnostics. Final-answer-only
	records never enter the canonical trajectory population.

	Outcome and provenance are independent. Human/reference demonstrations,
	recorded model rollouts, source evaluator verdicts and local reproduction are
	never collapsed into one “expert” flag. No trajectory is called optimal.

	Licenses and training restrictions are recorded even though this task prepares
	data rather than training a model. Evaluation-only sources are retained for
	baseline analysis and assigned `protected_evaluation`; they are ineligible for
	default SFT export.