Standardized dataset card (gen_dataset_cards.py)

README.md

@@ -1,4 +1,17 @@
 ---
+license: apache-2.0
+pretty_name: Noval-Corp Training Corpus (135M round)
+language:
+  - en
+task_categories:
+  - text-generation
+tags:
+  - continued-pretraining
+  - domain-adaptation
+  - synthetic
+  - fictional
+size_categories:
+  - 1K<n<10K
 dataset_info:
   features:
   - name: text
@@ -20,3 +33,41 @@ configs:
   - split: validation
     path: data/validation-*
 ---
+
+# Noval-Corp Training Corpus (135M round)
+
+**What's inside** — Plain-text training documents for the first (SmolLM2-135M) CPT round: Orbital Mining Corporation technical docs and Mars Express telemetry logs.
+
+**Where it comes from** — `orbital-mining-corpus` + `mars-express-telemetry-logs`, interleaved ~70/30 in favour of OMC.
+
+**How it was built / modified** — Interleaving only — no curation, dedup or replay (those arrived in later rounds). Validation is the untouched OMC test split. Superseded by `noval-corp-training-corpus-lfm2` and `asterion-training-corpus-gemma4`.
+
+## Overview
+
+- **Role**: CPT corpus — first (SmolLM2-135M) round
+- **Lineage**: `orbital-mining-corpus` + `mars-express-telemetry-logs` → **this** → `smollm2-135m-noval` (CPT)
+
+## Schema
+
+| Column | Meaning |
+|---|---|
+| `text` | one document per row (OMC doc or telemetry log) |
+
+## Splits
+
+| Split | Rows | Note |
+|---|---|---|
+| train | 8,272 | interleaved 30/70 mars/orbital |
+| validation | 45 | orbital-mining-corpus test split |
+
+## Used by
+
+- train.py (135M round) → atenareply/smollm2-135m-noval
+
+## Limitations
+
+- Fictional domain (OMC) — the corpus teaches an invented world.
+- No curation/dedup/replay: those improvements arrived with the LFM2 round.
+
+---
+_Part of the **noval-corp** fictional-domain training family (Orbital Mining Corporation / Mars Express; Asterion round for Gemma-4). Card generated by `noval-corp/scripts/gen_dataset_cards.py` — see `noval-corp/NOTE.md` for the full project log._