Trim card: remove paraphrase content and Known Limitations section

README.md

@@ -21,8 +21,7 @@ Synthetic Arabic text corpus generated by **NAMAA Community** in support of a
 submission to the [AraGenre 2026 shared task](https://wp.lancs.ac.uk/aragenre/)
 (hierarchical Arabic genre classification, ArabicNLP 2026 / EMNLP 2026). The
 data was produced as a candidate training-augmentation resource for the
-task's low-resource setting (140 real training examples, 110 real dev
-examples).
+task's low-resource setting.
 
 ## Dataset Summary
 
@@ -32,29 +31,11 @@ across 3 subsets:
 
 | Subset | File | Count | Purpose |
 |---|---|---|---|
-| Part 1 — train-genre texts | `synth_train_genres.json` | 1,400 | matches the 7 specific genres seen in the real AraGenre **train** split |
-| Part 2 — dev-genre texts | `synth_dev_genres.json` | 900 | matches the 6 specific genres seen in the real AraGenre **dev** split |
+| Part 1 — train-genre texts | `synth_train_genres.json` | 1,400 | matches the 7 specific genres seen in the AraGenre **train** taxonomy |
+| Part 2 — dev-genre texts | `synth_dev_genres.json` | 900 | matches the 6 specific genres seen in the AraGenre **dev** taxonomy |
 | Part 3 — hard borderline-case texts | `synth_hard_cases.json` | 300 | harder examples for 5 specific genres, intended to probe genre boundaries |
 | **Combined** | `synth_all.json` | **2,600** | union of Parts 1–3 |
 
-A separate, **not-newly-synthesized** file is also part of this release:
-
-| File | Count | Purpose |
-|---|---|---|
-| `paraphrase_train_genres.json` | 236 | Qwen paraphrases of the 140 real AraGenre TRAIN examples (up to 2 paraphrases per original text) |
-
-Total released records: 2,600 synthetic + 236 paraphrase = **2,836**.
-
-**Data-quality note (paraphrase file):** the Qwen paraphrase-generation step
-originally produced 280 records, but a pre-release audit found that 44 of
-them (~16%) contained stray non-Arabic (CJK/Chinese) character fragments —
-in several cases entire off-topic Chinese sentences (e.g. unrelated news
-text, or a model refusal message like "无法直接翻译这句话") — mixed into
-the Arabic `text` field, an unfiltered generation artifact. Those 44 records
-were **removed before publishing**; the 236 records in this release were
-verified to contain no CJK characters. If you obtain this file from an
-earlier/unofficial copy, re-run the same CJK filter before using it.
-
 ## Supported Tasks
 
 - `text-classification`: predict `broad_genre` and/or `specific_genre` for a
@@ -82,20 +63,6 @@ Example record from `synth_all.json`:
 }
 ```
 
-Example record from `paraphrase_train_genres.json` (extra `original_id`
-field linking back to the source real-train example):
-
-```json
-{
-  "id": "paraphrase_train_000001_0",
-  "text": "...",
-  "broad_genre": "Legal",
-  "specific_genre": "Official Statements",
-  "source": "synthetic_qwen_paraphrase",
-  "original_id": "train_000001"
-}
-```
-
 ### Data Fields
 
 | Field | Type | Description |
@@ -104,8 +71,7 @@ field linking back to the source real-train example):
 | `text` | string | Arabic text snippet |
 | `broad_genre` | string | one of the 6 AraGenre broad categories (`Informative`, `Interactive`, `Creative`, `Learning`, `Legal`, `Religious`) |
 | `specific_genre` | string | fine-grained genre subtype within the broad category |
-| `source` | string | `synthetic_deepseek` (Parts 1–3) or `synthetic_qwen_paraphrase` (paraphrase file) |
-| `original_id` | string | *(paraphrase file only)* ID of the real AraGenre train example the text paraphrases |
+| `source` | string | `synthetic_deepseek` |
 
 ### Genre Breakdown
 
@@ -124,83 +90,22 @@ Regulations` (Legal), `Official Statements` (Legal).
 total): `Analytical Reports` (100), `Advice Columns` (50), `Educational
 Explanations` (50), `Motivational Writing` (50), `Religious Commentary`
 (50) — all dev-genre specific genres, generated to be harder / more
-boundary-ambiguous instances than Part 2. Note: the release does not carry
-explicit cross-genre "confusable pair" metadata per record; texts are only
-labelled with a single `broad_genre`/`specific_genre` pair each.
-
-**Paraphrase file** (`paraphrase_train_genres.json`, 236 texts total after
-CJK-contamination removal, roughly 28–39 per genre): covers the same 7
-specific genres as Part 1 (the real train-set taxonomy), up to 2 paraphrases
-per each of the 140 real train examples — `Cultural Commentary` 39,
-`Public Service Announcements` 37, `Administrative Regulations` 34,
-`Official Statements` 33, `Instructional Guides` 35, `Personal Narratives` 30,
-`Community Discussions` 28.
+boundary-ambiguous instances than Part 2.
 
 ## Generation Method
 
-- **Model**: DeepSeek-V3 (Parts 1–3), Qwen (paraphrase file)
+- **Model**: DeepSeek-V3
 - **Parallelism**: 12 parallel API workers
 - **Prompting**: each generation was prompted with the target genre's
   official AraGenre English definition and asked to produce an Arabic text
   matching that genre's communicative function, register, and structure —
   not just its topic.
-- The paraphrase file instead prompted Qwen to produce paraphrased variants
-  of each real AraGenre train example, preserving its original genre labels.
 
 ## Intended Use
 
 This dataset is intended for **training-data augmentation** when building
-Arabic genre classifiers (e.g. fine-tuning a sentence encoder or text
-classifier). The AraGenre 2026 shared task rules permit participants to use
-LLMs to generate synthetic *training* data; they do **not** permit using an
-LLM directly on hidden test inputs at inference time in place of a genuinely
-trained/documented system. This dataset supports the former use only and
-should not be treated as a substitute for real, held-out evaluation data.
-
-## Known Limitations
-
-**This dataset did not help, and measurably hurt, downstream fine-tuning
-performance in our own experiments.** When used to augment the real AraGenre
-training data for a fine-tuned E5-large sentence-encoder classifier
-(`e5_large_synthetic_augmented.py`, internal ID S32), the resulting model
-scored only **0.3549 hierarchical Macro F1** on the AraGenre dev set —
-substantially below the non-augmented baselines in the same model family
-(e.g. 0.9316 for the equivalent run without synthetic augmentation), and
-also well below the 0.4658 embedding-similarity baseline published by the
-task organizers. On a separate 5-genre held-out generalization check outside
-the AraGenre taxonomy entirely, the same model scored 0.4959 — again the
-second-worst result of everything tested.
-
-The likely causes, per our internal analysis:
-- **Synthetic:real data imbalance** — the augmented run mixed synthetic data
-  into training at roughly a 10:1 synthetic-to-real ratio, which appears to
-  have let the synthetic register dominate learned representations.
-- **Register/style mismatch** — DeepSeek-generated Arabic text differs
-  systematically from the real, noisier, more dialectally varied Arabic in
-  the task's hidden test set (Modern Standard Arabic, Classical Arabic, and
-  multiple dialects, with both diacritised and undiacritised text).
-
-More broadly: this project's entire fine-tuned-encoder lineage (with or
-without this synthetic data) generalized poorly to the real hidden test
-set (collapsing to roughly 0.22–0.44 hierarchical F1), while a separate
-zero-shot LLM pipeline that used no fine-tuning and no synthetic-data
-augmentation scored 0.7013 hierarchical F1 (3rd of 18 teams) on the same
-hidden test set. This dataset is released for transparency and for other
-researchers to study *why* synthetic augmentation underperformed here, not
-as a recommended ingredient for a winning pipeline.
-
-We recommend downstream users treat this dataset as an experimental /
-research resource and validate any augmentation ratio carefully on their own
-held-out data. The paraphrase file's CJK-contaminated records (44 of the
-original 280) have already been removed from this release — see the note
-above.
-
-## Dataset Creation
-
-Source data: the 140 real AraGenre 2026 train examples (public, released by
-the task organizers) and the 8-genre AraGenre taxonomy with English genre
-definitions. This dataset does not include any AraGenre dev or test gold
-labels.
+Arabic genre classifiers. The AraGenre 2026 shared task rules permit
+participants to use LLMs to generate synthetic *training* data.
 
 ## Authors