Upload README.md with huggingface_hub

README.md

@@ -193,6 +193,55 @@ The glossary can be used for:
 - Orthographic variation may occur due to non-standardized spelling conventions in Nigerian Pidgin.
 - Certain educational concepts may have multiple valid Pidgin renderings depending on context.
 
+---
+
+# Glossary-Augmented Dataset (Eng-PidginEdu v2)
+
+`Eng-PidginEdu_glossary_augmented.csv`, also included in this repository,
+is the parallel corpus above with the glossary merged directly into the
+Pidgin translations: matched academic terms get their Pidgin meaning
+inserted inline, in brackets, immediately after the term.
+
+```
+en:            The ability for educators and students to adopt the
+               materials for free.
+pcm:           Teachers and students fit get the materials for free.
+pcm_augmented: Teachers and students fit get the materials (tins wey
+               dem use for teaching) for free.
+```
+
+This glossary-augmented file -- not the plain parallel corpus -- is the
+actual training target used to fine-tune all 14 models in the
+[Eng-PidginEdu benchmark](https://github.com/shashacode/Eng-PidginEdu-Benchmark),
+and is the source of this project's novel glossary-accuracy evaluation
+metric.
+
+## Columns
+
+In addition to `en`, `pcm`, and `Subject` above:
+
+| Column | Description |
+|---|---|
+| `pcm_augmented` | The Pidgin translation with inline glosses inserted |
+| `glossed_terms` | Semicolon-separated list of which terms were glossed in this row |
+| `n_glosses` | Number of glosses inserted in this row |
+
+## Augmentation statistics
+
+| Metric | Value |
+|---|---:|
+| Rows with at least one gloss | 16,114 / 26,232 (61.4%) |
+| Total glosses inserted | 28,243 |
+| Distinct terms glossed | 2,473 |
+
+## Reproducing it
+
+The exact preprocessing script that produces this file from the raw
+corpus and the standalone glossary table above is `glossary_augment.py`,
+published in the
+[benchmark's GitHub repository](https://github.com/shashacode/Eng-PidginEdu-Benchmark) --
+verified to reproduce this file exactly, byte-for-byte, across all
+26,232 rows.
 
 ---