alexneakameni/ZSHOT-HARDSET-v2

generate-gliznet-data

Generation pipeline for ZSHOT-HARDSET-v2 and ZSHOT-HARDSET-Polarity — hard-negative zero-shot text classification benchmarks.

---

Overview

This repository contains the generation pipelines for two complementary zero-shot text classification datasets. Unlike standard ZSC datasets that rely on surface-level topic matching, these datasets target semantic understanding: labels describe meaning, intent, rhetorical stance, and epistemic function — not keywords.

ZSHOT-HARDSET-v2 (Wikipedia)

Each example contains:

• text — an original short passage (2–3 sentences) written in a specific genre/register
• labels — 3–5 labels that are semantically true for this text
• not_labels — 3–5 labels that are plausible but wrong: they would fool a naive classifier but are clearly false to a careful reader

ZSHOT-HARDSET-Polarity (Amazon Reviews)

Each example contains:

• text — an Amazon product review (title + content)
• sentiment — the ground-truth sentiment (positive / negative)
• labels — 50 labels that are semantically true for this review
• not_labels — 50 labels that are plausible but wrong

The key difference from the Wikipedia variant is the scale of labels: 50 positive and 50 negative labels per example, requiring fine-grained semantic distinctions about rhetorical stance, emotional register, argumentative moves, writing style, and epistemic signals.

---

Generation Pipelines

Wikipedia Pipeline (ZSHOT-HARDSET-v2)

1. Source: Wikipedia (streaming)

Articles are sampled from the wikimedia/wikipedia 20231101.en split via HuggingFace Datasets in streaming mode with a large shuffle buffer. Wikipedia provides broad topical diversity while keeping content grounded in factual reality.

2. Genre injection (70 registers)

Each Wikipedia article is paired with a randomly sampled genre from a vocabulary of 70 text registers, spanning:

• Factual: encyclopedia entry, academic abstract, empirical paper finding, patent claim …
• Journalistic: news article lede, investigative journalism, tabloid lede …
• Conversational: Reddit post, forum Q&A, podcast transcript, chat/SMS …
• Creative / narrative: myth retelling, documentary narration, song lyrics, children's book …
• Institutional: legal document, press release, diplomatic statement, parliamentary debate …
• Personal: diary entry, eulogy, oral history, customer complaint …
• And many more

This forces the model to rewrite factual content into diverse registers, producing texts that vary in rhetorical stance, epistemic function, and social context — not just topic.

3. LLM generation (vLLM)

Bundles of 5 texts + 18 shared semantic labels are generated per article using a local LLM served via vLLM. The prompt enforces three hard constraints:

• Labels must capture meaning and intent, not surface vocabulary
• Every label must appear as a positive in at least one text and a negative in at least one different text (cross-role constraint)
• not_labels must require actually reading the text to rule out — not obviously unrelated

Model: Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled (reasoning-distilled, 27B)

4. Post-processing

• Bundles are parsed and validated; malformed JSON is discarded
• Rows with the same text are merged: labels and not_labels are unioned, and any intersection is removed from not_labels

5. Train / Test split

The dataset uses a label-based split to encourage zero-shot evaluation:

1. A small fraction of positive labels is randomly selected as held-out test labels
2. Rows containing at least one held-out label in labels are assigned to test
3. Rows where any held-out label appears in either labels or not_labels are excluded from train

This yields ~30% novel labels in the test set (labels never seen during training), forcing the model to generalise to unseen semantic categories.

Amazon Polarity Pipeline (ZSHOT-HARDSET-Polarity)

1. Source: Amazon Polarity

Reviews are streamed from fancyzhx/amazon_polarity via HuggingFace Datasets. Each review includes a title, content, and binary sentiment label.

2. LLM annotation

Each review is annotated by an LLM that generates 50 positive labels and 50 negative labels per review. Labels describe:

• Rhetorical stance (what claim the reviewer is making and how)
• Emotional register (feeling the text conveys or appeals to)
• Argumentative moves (evidence or reasoning structure)
• Implied relationship to the product (expert, casual buyer, gift buyer…)
• Writing style markers (colloquial, formal, hyperbolic, hedged…)
• Epistemic signals (certainty, doubt, recommendation, warning…)
• Discourse function (introduces, contrasts, exemplifies, concedes…)

Labels are short snake_case phrases (2–5 words) and must not include product names, brands, category keywords, or generic sentiment words.

3. Post-processing

• Parsed and validated; malformed JSON or under-generated annotations (< 5 labels) are discarded
• Any label appearing in both labels and not_labels is removed from both lists

---

Repository Structure

generate-gliznet-data/
├── main.py                  # Wikipedia pipeline: stream, generate, write JSONL
├── annotate_amazon.py       # Amazon Polarity pipeline: annotate reviews
├── launch.sh                # Launch command for Wikipedia generation
├── launch_amazon.sh         # Launch command for Amazon annotation (local vLLM)
├── launch_amazon_openrouter.sh  # Launch command for Amazon annotation (OpenRouter)
├── src/
│   ├── prompt.py            # System prompt for Wikipedia generation
│   ├── prompt_amazon.py     # System prompt for Amazon annotation
│   ├── genres.py            # 70 text genre definitions
│   └── config.py            # Token budget and generation constants
├── notebooks/
│   └── CreateDataset.ipynb  # Load JSONL → merge → split → push to HF Hub
└── data/                    # Generated JSONL files (gitignored)

---

Quickstart

pip install vllm datasets tqdm llm-output-parser

python main.py \
    --output_path data/wikipedia_synthetic.jsonl \
    --model Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled \
    --num_examples 64000 \
    --batch_size 16 \
    --tensor_parallel_size 2 \
    --max_tokens 4096 \
    --temperature 0.9 \
    --seed 42 \
    --shuffle_buffer 50000

---

Links

• ZSHOT-HARDSET-v2 (Wikipedia): alexneakameni/ZSHOT-HARDSET-v2
• ZSHOT-HARDSET-Polarity (Amazon): alexneakameni/ZSHOT-HARDSET-Polarity
• Generation code: KameniAlexNea/generate-gliznet-data

---

License

See LICENSE.