DATA & TRAINING
Configured methods are not completed runs.
This record distinguishes dataset provenance, configured targets, smoke-test evidence and absent production or training artifacts across the Quantum model line.
PRIMARY DATASET
epfml/FineWeb2-HQ
deu_Latn
A German-language subset derived from public web material distributed through FineWeb2-HQ.
The dataset card identifies ODC-By 1.0 and Common Crawl terms. Those terms do not automatically clear rights in every source document or license model weights, code, tokenizers or visual assets.
Inspect dataset cardquantum-1-pilot
- Source revision
- Configured as mutable `main`
- Configured target
- 100M training, 1M validation and 1M test tokens at 512-token context.
- Public observation
- The model card reports approximately 100M training tokens; no final data manifest is public.
quantum-1.6-pilot
- Source revision
- Configured as mutable `main`
- Configured target
- 500M new training, 2M validation and 2M test tokens at 512-token context.
- Public observation
- The release card reports 500M additional tokens; final document counts, split hashes and overlap statistics are not public.
quantum-1-echelon
- Source revision
- c0c06e94fd3a44ae9e802b2b0fc533817601eb5e
- Configured target
- 8B training, 10M validation and 10M test tokens at 2,048-token context.
- Public observation
- A smoke run saw 5,001 documents, accepted 1,559 and produced 1,380,886 packed tokens. The production run had not started.
PIPELINE CONTROLS
Documented controls.
- Language, script, length, quality, URL, digit and symbol filtering
- Deterministic split assignment and configured seeds
- Exact-document controls and reliance on upstream MinHash deduplication in the Echelon pipeline
- Packed token storage with checkpoint and resume handling
KNOWN LIMITATIONS
Evidence still missing.
- Exact fingerprints do not remove every near-duplicate or semantically overlapping document.
- Public web data may retain personal, harmful, copyrighted or otherwise sensitive material.
- The pilot configurations do not resolve their mutable dataset revision in a final manifest.
- No public record links final accepted documents and split hashes to either pilot GGUF.
- No production Echelon dataset or completed Echelon training run is public.
Questions about provenance, privacy, rights or removal should be sent to the public project email with the affected source and enough information to identify it. Contact: cikemgil@rappidai-research.com
DOCUMENTATION DIRECTORY
Continue through the evidence.
Every section keeps public evidence, configured targets and missing information visibly separate.