Applied AI · Training data · Python
Training data,
released with proof.
- Declared stages
- 12
- Automated tests
- 242
- Release gates
- 7
- Artifact identity
- SHA-256
Can a dataset explain why it deserves to be released?
Training-data workflows often preserve the final JSONL while losing the decisions that produced it: source permissions, rejected records, duplicate handling, benchmark overlap, split isolation and the exact configuration used.
Forge treats the dataset as a release rather than a folder. One typed pipeline core drives the command line, Python API, tests and hosted demonstration, while each stage emits inspectable evidence and content-derived identities.
Four boundaries from source to verified artifact.
- 01Govern
Fingerprint sources, declare permitted use and reject or redact detected structured identifiers.
- 02Curate
Generate typed examples, score quality and quarantine exact, fuzzy or semantic duplicates.
- 03Isolate
Check benchmark contamination and partition entire documents without source leakage.
- 04Verify
Enforce release gates, hash every artifact and independently verify the resulting manifest.
Quality claims need inspectable boundaries.
The implementation separates controls that execute from model-quality evidence that has not yet been established.
Source rights
Candidate releases require a declared source-rights record and permitted training use for every loaded document.
Privacy boundaries
Source and generated records pass deterministic structured-identifier checks before they can enter a release.
Duplicate control
Exact, MinHash, Jaccard and optional embedding checks quarantine duplicates with machine-readable reason codes.
Contamination gate
A supplied benchmark is checked independently and can block a candidate release when overlap crosses the configured threshold.
Source isolation
Whole documents are assigned to one partition, and the split manifest refuses document or chunk overlap across train and test.
Artifact integrity
The release manifest hashes its inputs and outputs, verifies counts and refuses modified or incomplete artifacts.
One pipeline core, several inspection surfaces.
What Forge proves today.
01Pipeline executionThe hosted smoke release invokes the real workflow and streams all twelve declared stages.
02Release integritySeven gates check stage success, contamination, split isolation, counts and artifact hashes before release.
03Source-safe splitsRecords remain grouped by document identity, and overlap checks are recorded in the split manifest.
04Failure recoveryResumed stages require matching contracts, input hashes, configuration and verified output hashes.
05Honest evaluationThe included calibration path proves the evaluation controls execute; fixture labels are not presented as genuine human evidence.
A verified smoke release is not a model-quality result.
Public smoke runs disable semantic embeddings and use deliberately small synthetic fixtures. They establish pipeline execution, artifact integrity and declared control behavior, not production model quality or complete sensitive-data detection.
The blind evaluation system is implemented, but the required genuine human annotation collection has not been completed. Source-rights records are operator declarations, hosted state is intentionally in memory and the local fine-tuning path is MLX-specific.