Upload README.md with huggingface_hub

README.md

@@ -1,89 +1,261 @@
 ---
 task_categories:
 - text-generation
-- question-answering
+- information-extraction
 language:
 - en
 size_categories:
 - n<1K
+tags:
+- web-scraping
+- html-extraction
+- structured-data
+- synthetic-data
 ---
 
-# CrawlerLM: HTML-to-JSON Dataset
+# CrawlerLM: HTML Fragment to Structured JSON
 
-A synthetic dataset for training language models to extract structured JSON data from raw HTML.
+A synthetic dataset for training language models to extract structured JSON from HTML fragments across multiple schema types.
 
 ## Dataset Description
 
-This dataset contains HTML pages paired with their structured JSON representations, designed for fine-tuning small language models for web scraping and information extraction tasks.
+This dataset contains HTML fragments paired with structured JSON annotations across three schema types: **recipes**, **job postings**, and **events**. It's designed for fine-tuning small language models to perform domain-specific information extraction from messy, real-world HTML.
 
-### Dataset Structure
+### Key Features
 
-Each example contains:
-- `example_html`: Raw HTML content from real web pages
-- `expected_json`: Structured extraction with fields:
-  - `url`: Page URL
-  - `title`: Page title
-  - `text`: Main text content
-  - `author`: Author name (or null)
-  - `published_date`: Publication date (or null)
-  - `image`: Main image URL
-  - `favicon`: Favicon URL
-  - `id`: Unique identifier
+- **60 manually annotated base examples** from diverse web sources
+- **3 schema types** with domain-specific fields
+- **Synthetic augmentation** to 447+ training examples with realistic HTML variations
+- **Two configurations**: raw (HTML→JSON) and chat (instruction format)
+- **Token-filtered** (all examples ≤24K tokens)
 
-### Data Splits
+## Configurations
 
-- **Train**: 450 synthetic variations
-- **Test**: 50 synthetic variations
+### `raw` Configuration
 
-### Data Sources
+HTML fragment to JSON extraction format.
 
-- Base HTML samples from Common Crawl
-- Structured extractions via Exa API
-- Synthetic variations generated programmatically
+**Fields**:
+- `example_html` (string): Raw HTML fragment
+- `expected_json` (dict): Structured extraction with schema-specific fields
 
-### Use Cases
+**Example**:
+```python
+{
+  "example_html": "<div class=\"recipe-card\">...</div>",
+  "expected_json": {
+    "type": "recipe",
+    "title": "Best Ever Macaroni Cheese",
+    "ingredients": ["500g macaroni", "200g cheddar", ...],
+    "instructions": ["Boil pasta", "Make sauce", ...],
+    "prep_time": "10 mins",
+    "cook_time": "20 mins",
+    ...
+  }
+}
+```
+
+**Splits**:
+- Train: 400 examples (augmented from 48 base examples)
+- Validation: 50 examples (augmented from 6 base examples)
+- Test: 6 examples (pristine, no augmentation)
+
+### `chat` Configuration
+
+Instruction-tuning format for training chat models.
+
+**Fields**:
+- `messages` (list): Conversational format with user/assistant roles
+
+**Example**:
+```python
+{
+  "messages": [
+    {
+      "role": "user",
+      "content": "Extract structured data from the following HTML and return it as JSON.\n\nHTML:\n<div>...</div>"
+    },
+    {
+      "role": "assistant",
+      "content": "{\"type\": \"recipe\", \"title\": \"...\", ...}"
+    }
+  ]
+}
+```
+
+**Splits**:
+- Train: 391 examples (9 filtered out for exceeding token limit)
+- Validation: 50 examples
+- Test: 6 examples
+
+## Schema Types
+
+### Recipe (`type: "recipe"`)
+
+**Fields**: `type`, `title`, `description`, `ingredients`, `instructions`, `prep_time`, `cook_time`, `total_time`, `servings`, `cuisine`, `difficulty`, `rating`, `author`, `image_url`, `video_url`, `source_url`, `published_date`
+
+**Use case**: Extracting recipe data from food blogs, cooking sites
+
+**Example sources**: BBC Good Food, AllRecipes, Serious Eats
 
-- Fine-tuning small models for web scraping
-- Training HTML-to-JSON extraction models
-- Benchmarking structured data extraction
+### Job Posting (`type: "job_posting"`)
+
+**Fields**: `type`, `title`, `company`, `location`, `compensation`, `benefits`, `mode_of_work`, `job_type`, `experience_level`, `requirements`, `responsibilities`, `description`, `application_url`, `company_logo`, `source_url`
+
+**Use case**: Parsing job listings from career pages, job boards
+
+**Example sources**: Greenhouse, Lever, LinkedIn Jobs
+
+### Event (`type: "event"`)
+
+**Fields**: `type`, `title`, `description`, `datetime`, `end_datetime`, `location`, `venue`, `organizer`, `price`, `registration_url`, `image_url`, `category`, `tags`, `source_url`
+
+**Use case**: Extracting event details from event listings, calendars
+
+**Example sources**: Eventbrite, Meetup, local event pages
+
+## Data Collection Process
+
+1. **Manual Annotation**: 61 HTML fragments manually annotated using custom Chrome extension
+2. **Quality Filtering**: Removed 1 example exceeding 24K token limit (60 examples remaining)
+3. **Stratified Split**: 80/10/10 split by schema type (48 train / 6 val / 6 test base examples)
+4. **Synthetic Augmentation**:
+   - Train: ~8 variations per base example (400 total)
+   - Val: ~8 variations per base example (50 total)
+   - Test: No augmentation (6 pristine examples)
+5. **Chat Conversion**: Convert to instruction-tuning format with token filtering
+
+### Augmentation Strategies
+
+- **Structural variations**: Wrapper divs, nesting depth changes
+- **Attribute noise**: Random classes, IDs, data-* attributes
+- **Template variations**: Semantically equivalent tags (div ↔ section)
+- **HTML comments**: Developer comments injection
+- **Whitespace variations**: Minified vs. prettified formatting
+
+All augmentations preserve semantic content and ensure `expected_json` remains unchanged.
 
 ## Usage
 
+### Load Raw Configuration
+
 ```python
 from datasets import load_dataset
 
-dataset = load_dataset("espsluar/crawlerlm-html-to-json")
+# Load raw HTML→JSON format
+dataset = load_dataset("espsluar/crawlerlm-html-to-json", "raw")
 
-# Access splits
 train_data = dataset["train"]
+val_data = dataset["validation"]
 test_data = dataset["test"]
 
-# Example
+# Inspect example
 example = train_data[0]
+print(f"Schema type: {example['expected_json']['type']}")
 print(f"HTML length: {len(example['example_html'])} chars")
 print(f"Title: {example['expected_json']['title']}")
 ```
 
-## Dataset Creation
+### Load Chat Configuration
 
-Generated using the CrawlerLM pipeline:
-1. Sample diverse URLs from Common Crawl
-2. Filter for quality (SPA detection, content scoring)
-3. Extract structured data via Exa API
-4. Generate synthetic variations (wrappers, noise, perturbations)
+```python
+from datasets import load_dataset
 
-## License
+# Load chat format for instruction tuning
+dataset = load_dataset("espsluar/crawlerlm-html-to-json", "chat")
 
-MIT
+train_data = dataset["train"]
+
+# Inspect example
+example = train_data[0]
+print(f"User prompt: {example['messages'][0]['content'][:100]}...")
+print(f"Assistant response: {example['messages'][1]['content'][:100]}...")
+```
+
+### Filter by Schema Type
+
+```python
+# Filter for only recipes
+recipes = dataset["train"].filter(
+    lambda x: '"type": "recipe"' in x["messages"][1]["content"]
+)
+
+print(f"Recipe examples: {len(recipes)}")
+```
+
+## Dataset Statistics
+
+| Split | Raw Examples | Chat Examples | Schema Distribution |
+|-------|--------------|---------------|---------------------|
+| Train | 400 | 391 | ~133 recipe, ~150 job_posting, ~117 event |
+| Validation | 50 | 50 | ~17 recipe, ~17 job_posting, ~16 event |
+| Test | 6 | 6 | 2 recipe, 2 job_posting, 2 event |
+
+**Schema Distribution** (base examples before augmentation):
+- Recipe: 19 examples (31.7%)
+- Job Posting: 22 examples (36.7%)
+- Event: 19 examples (31.7%)
+
+## Intended Use
+
+### Primary Use Cases
+
+- Fine-tuning small language models (0.5B-7B parameters) for HTML extraction
+- Training domain-specific web scrapers
+- Benchmarking structured data extraction performance
+- Teaching models to handle messy, real-world HTML
+
+### Out of Scope
+
+- Full webpage extraction (this dataset focuses on **fragments**, not entire pages)
+- Single-field extraction (schemas have 10-17 fields each)
+- Non-English content
+- Dynamic/JavaScript-rendered content
+
+## Limitations
+
+- **Limited schema types**: Only 3 schema types (recipe, job_posting, event)
+- **English only**: All examples are from English-language websites
+- **Static HTML**: No JavaScript-rendered or dynamic content
+- **Token limit**: All examples ≤24K tokens (may not represent very long pages)
+- **Augmentation artifacts**: Synthetic variations may not perfectly match real-world HTML diversity
+
+## Ethical Considerations
+
+- **Web scraping**: This dataset is intended for educational and research purposes. Users should respect robots.txt and website terms of service when deploying trained models.
+- **Data sources**: All HTML fragments are from publicly accessible websites
+- **Privacy**: No personally identifiable information (PII) is intentionally included
 
 ## Citation
 
 ```bibtex
 @misc{crawlerlm2025,
   author = {Jack Luar},
-  title = {CrawlerLM: HTML-to-JSON Dataset},
+  title = {CrawlerLM: HTML Fragment to Structured JSON},
   year = {2025},
   publisher = {HuggingFace},
   howpublished = {\url{https://huggingface.co/datasets/espsluar/crawlerlm-html-to-json}}
 }
 ```
+
+## License
+
+MIT
+
+## Dataset Creation
+
+**Tooling**: Custom Chrome extension for manual annotation ([github.com/espsluar/c4ai-crawlerlm](https://github.com/espsluar/c4ai-crawlerlm))
+
+**Pipeline**:
+1. Manual HTML fragment selection and annotation
+2. Schema-specific field extraction
+3. Quality filtering (token limits, validation)
+4. Stratified train/val/test split
+5. Synthetic augmentation (structural, attribute, whitespace variations)
+6. Chat format conversion with instruction templates
+
+**Quality Control**:
+- Manual review of all base annotations
+- Token count validation (≤24K per example)
+- Schema validation (required fields, types)
+- Stratified sampling to ensure balanced schema distribution