sphita/Intel-WebCorpus-forms

💻 Intel WebCorpus Forms (Enterprise Hardware Q&A)

This dataset is a massive, high-fidelity archive of 176,472 technical troubleshooting discussions (containing nearly 1 million individual messages) scraped from the official Intel Community Forums. It has been meticulously engineered for Large Language Model (LLM) training. Instead of a raw, messy dump of isolated posts, the data has been reconstructed into chronological conversation threads, noise-filtered, deduplicated, and packed into optimized Snappy-compressed Parquet files.

📊 Dataset Summary

Metric	Count
Total Discussions (Threads)	176,472
Total Unique Messages	939,959
Average Messages per Thread	5.3
Threads with Accepted Solutions	20,935 (11.9%)
Total Parquet Chunks	1,765
Total Size (Compressed)	~360 MB

🛠️ Data Engineering & Quality

This dataset was built using a custom Rust-based async crawler designed to bypass Enterprise WAFs and handle extreme server-side rate limits using exponential backoff. The raw data pipeline included:

1. Thread Reconstruction: Over 900,000 flat messages were grouped by their topic_id to reconstruct the exact back-and-forth conversational flow of human troubleshooting.
2. Noise Reduction: "Dead-weight" messages (under 5 words, e.g., "bump", "+1", "thanks") were automatically filtered out of the replies to increase the signal-to-noise ratio.
3. Deduplication: Overlapping chunks were strictly deduplicated by msg_id to ensure absolute uniqueness.
4. HTML Sanitization: Raw HTML bodies were stripped into clean text.

🗂️ Top Hardware Categories (Boards)

The discussions cover deep technical support for Intel's entire hardware and software ecosystem. The top 10 categories are:

Board / Category	Discussions
fortran-compiler	28,479
graphics (Intel Arc & Iris)	23,721
processors (Core & Xeon)	17,436
software-archive	16,556
wireless	9,006
oneapi-math-kernel-library	7,198
integrated-performance-primitives	6,675
distribution-openvino-toolkit	6,621
ethernet-products	5,942
server-products	5,311

📝 Schema

The dataset is provided in parquet format. Each row represents a complete discussion thread.

Column	Type	Description
discussion_id	string	The unique ID of the forum topic/thread.
subject	string	The title or subject of the discussion.
board	string	The specific Intel hardware/software category.
post_time	string	ISO 8601 timestamp of the original question.
has_solution	boolean	True if an Intel engineer or community member provided an accepted solution.
message_count	int64	The total number of valid messages in the thread.
messages_json	string	A JSON-encoded string containing the full list of messages in chronological order.

🔍 Inside messages_json

If you parse the messages_json column, you get a list of dictionary objects representing the conversation:

[
  {
    "msg_id": "12345",
    "depth": 0,
    "author": "user_123",
    "post_time": "2023-01-01T12:00:00Z",
    "is_solution": false,
    "body": "I am getting Error 504 on my Xeon server...",
    "kudos": 2
  },
  {
    "msg_id": "12346",
    "depth": 1,
    "author": "Intel_Support",
    "post_time": "2023-01-01T12:15:00Z",
    "is_solution": true,
    "body": "Please update your BIOS to version X...",
    "kudos": 5
  }
]

🚀 Use Cases

• Hardware Troubleshooting LLMs: Train models that understand complex server, networking, and processor error codes.
• Reasoning (Chain-of-Thought): Train models on multi-turn troubleshooting processes, where users try solutions, fail, report back, and eventually succeed.
• RAG (Retrieval-Augmented Generation): Build enterprise IT chatbots.

🔄 Updates

This dataset is maintained via an automated Delta Sync Pipeline. New discussions are fetched weekly, sorted, and pushed as new chunk files without overwriting the existing archive.