---
pretty_name: Italian Culture Midtraining v2 for Zagreus
language:
- it
license: other
task_categories:
- text-generation
size_categories:
- 10K<n<100K
tags:
- continued-pretraining
- italian-culture
- decontaminated
- provenance
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
  - split: validation
    path: data/validation-*.parquet
---

# Italian Culture Midtraining v2

Private, viewer-friendly continued-pretraining corpus for
`mii-llm/zagreus-0.4B-ita`. Repository: `FinancialSupport/italian-culture-midtrain-v2`.

| split | rows | words | target-tokenizer tokens (EOS included) |
|---|---:|---:|---:|
| train | 84,031 | 23,357,466 | 48,463,711 |
| validation | 780 | 200,197 | 422,661 |

## Viewer schema

`text` is the exact model-visible text. `title`, `category`, `categories`,
`source`, `source_url`, `license`, revision/document IDs, word/token counts,
and normalized hashes are separate columns for filtering and review.

## Merge and deduplication

- `FinancialSupport/test_cpt` was audited in full. It is derived from the same
  raw sources already re-ingested or superseded by v2, so zero legacy rows were
  added; adding it would duplicate content under different chunk boundaries.
- Five legacy Pinocchio rows were rejected by the owner-enforced source ban.
- `FinancialSupport/zagreus-cpt-italic-packed` is the tokenized form of that
  legacy corpus and was not decoded/re-added.
- SFT, DPO, OPD and eval repositories are different objectives and were not
  mixed into language-model pretraining data.
- The ADN/articles/web corpora were not added: they are broad news corpora, and
  some cards do not establish underlying-content licensing/provenance suitable
  for this release.

See `reports/merge_audit.json` for row-level decision counts.

## Safety and quality gates

- Exact 9-gram exclusion against question and option text for all 10,000 ITALIC
  items, applied to the final text visible to the model.
- Document-level deterministic train/validation split.
- Exact normalized SHA-256 deduplication and source-aware SimHash near-dedup.
- Remaining ITALIC overlap, split document/hash overlap, malformed HTML,
  replacement characters, and missing required metadata: **zero**.

## Sources and licenses

Mixed-source dataset: Wikipedia/Wikimedia projects, Wikidata, ArCo (Ministero
della Cultura), Kaikki/Wiktionary, UD Italian-PUD, FineWeb-2 replay, public-domain
texts and open manuals. The exact license and source URL are retained per row.
Because obligations differ (CC BY-SA, GFDL, CC0, ODC-By and public domain), the
Hub metadata uses `license: other`; review the per-row field before redistribution.

## Recommended use

Start from the **base** Zagreus model, run one CPT pass at sequence length 1024
with fp32 master weights + bf16 autocast, then re-run the short-format SFT. Do
not concatenate the legacy packed corpus or repeat this dataset for many epochs.