| --- |
| license: other |
| task_categories: |
| - text-generation |
| - language-modeling |
| language: |
| - en |
| tags: |
| - mbg |
| - model-bahasa-garuda |
| - english |
| - causal |
| - spatial |
| - temporal |
| - trinity-mirror |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: MBG-1.0.parquet |
| --- |
| |
| # MBG 1.0 — Dataset |
|
|
| *deepRcurs Labs / `@deeprcurs` · author: Mzed Imamkh / `@mzedimamkh`* |
|
|
| English-only corpora for the **MBG 1.0 (Model Bahasa Garuda)** project. This is the |
| external **dataset** archive; the training/evaluation code lives in the workspace |
| snapshot (the "controller"), and the model weights live in the **model** repo |
| `deeprcurs/MBG-1.0`. |
|
|
| ## Contents |
|
|
| | File | Rows | Source | Purpose | |
| |---|---|---|---| |
| | `MBG-1.0.parquet` | 326,080 | generated corpora | main train/val corpus (domain-tagged) | |
| | `probe_challenge.parquet` | 96 | `probe_challenge.txt` | Trinity-Mirror probe->MoE disambiguation benchmark | |
| | `english_L1.txt` | 200,000 | generator | L1 rung training corpus (17M) | |
| | `english_v3.txt` | 120,000 | `gen_corpus_v3.py` | quality corpus v3 (combinatorial, high diversity) | |
| | `english_v2.txt` | 120,000 | `gen_corpus_v2.py` | quality corpus v2 (template-based; superseded by v3) | |
| | `english_gen.txt` | 6,000 | generator | general generated corpus | |
| | `english_tiny.txt` | 80 | generator | L0 smoke corpus | |
| | `probe_challenge.txt` | 96 | generator | disambiguation benchmark raw | |
|
|
| All text is **English-only** (project rule). Generated, template-based English |
| sentences across the four Trinity-Mirror domains. **Corpus v3** is the current |
| quality focus: it keeps the four probe domains but moves from hand-listed |
| templates to combinatorial composition, giving far higher lexical and |
| structural diversity (see `source/gen_corpus_v3.py` in the model repo). |
|
|
| ## Columns |
|
|
| `MBG-1.0.parquet` — **`text`**, **`domain`** (`CAUSAL`, `SPATIAL`, `TEMPORAL`, |
| `GENERAL`), **`source`** (which generator), **`split`** (`train`/`val`, seeded & |
| reproducible, ~10% val stratified by source+domain). |
|
|
| `probe_challenge.parquet` — **`text`**, **`source`**, **`task`** (`probe_challenge`). |
|
|
| ## Usage (Python) |
|
|
| ```python |
| import pandas as pd |
| df = pd.read_parquet("hf://datasets/deeprcurs/MBG-1.0-data/MBG-1.0.parquet") |
| print(df.head(), df.domain.value_counts()) |
| ``` |
|
|
| ## Reproducibility |
|
|
| The split is produced with a fixed seed (default `0`). The generator and all |
| processing scripts are in the MBG-1.0 workspace snapshot (the controller repo); |
| regenerating the corpora from the same seed reproduces identical data. |
|
|
| ## License |
|
|
| **Dual License (custom, source-available)** — same terms as the model |
| (`deeprcurs/MBG-1.0/LICENSE.md`). See the model card for the full text. |
|
|
| - **Personal / Non-Commercial Use:** free, for private experimentation and |
| research. **Re-branding** and claiming the work as your own are **prohibited**. |
| - **Commercial Use / Re-branding / derivative redistribution:** requires the |
| author's **prior written permission** (commercial license fee applies). |
|
|
| This applies to the corpus and dataset. It is **not** an open-source/CC license; |
| it intentionally restricts redistribution and use. |
|
|
