| {"label": "broken_html_dump", "text": "<html><body><script>var x=undefined;throw new Error('Cannot read properties of null');</script><div class=\"error\">500 Internal Server Error</div><p>nginx/1.18.0</p></body></html>"} | |
| {"label": "stack_trace", "text": "Traceback (most recent call last):\n File \"/app/main.py\", line 42, in <module>\n result = process(data)\n File \"/app/utils.py\", line 17, in process\n return json.loads(response.text)\njson.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0)"} | |
| {"label": "corrupted_binary", "text": "\u0000\u0001\u0002\u0003\u0004\u0005\u0006\u0007\b\t\n\u000b\f\r\u000e\u000f\u00ff\u00fe\u00fd\u00fc\u00fb\u00fa\u00f9\u00f8\u00f7\u00f6\u00f5\u00f4\u00f3\u00f2\u00f1\u00f0"} | |
| {"label": "spam_ads", "text": "BUY NOW!!! LIMITED TIME OFF!!! Click here to claim your FREE iPhone 15 Pro Max! No credit card required! Act fast, offer expires in 24 hours! $$$ $$$ $$$"} | |
| {"label": "casino_spam", "text": "\ud83c\udfb0\ud83c\udfb0\ud83c\udfb0 WELCOME TO ROYAL CASINO \ud83c\udfb0\ud83c\udfb0\ud83c\udfb0 Play now and get 500% BONUS on your first deposit! Trusted by millions of players worldwide. 24/7 live support. Sign up today and start winning big! www.royalcasino.example.com"} | |
| {"label": "keyword_stuffing", "text": "loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan loan"} | |
| {"label": "auto_generated_filler", "text": "In today's fast-paced world, it is important to stay informed and connected. With the ever-changing landscape of technology, we must adapt to new tools and platforms. This article will explore the many ways in which you can improve your daily routine by leveraging the power of modern software solutions. Stay tuned for more updates."} | |
| {"label": "thin_content", "text": "Python is a programming language. It is popular. Many people use it. You can write code in Python. Python is easy to learn."} | |
| {"label": "listicle_filler", "text": "Top 10 Tips for Success: 1. Wake up early. 2. Drink water. 3. Exercise. 4. Eat healthy. 5. Read books. 6. Meditate. 7. Set goals. 8. Track progress. 9. Stay consistent. 10. Celebrate wins."} | |
| {"label": "basic_tutorial", "text": "To install Python on your computer, go to python.org and download the latest version. Run the installer and make sure to check the box that says 'Add Python to PATH'. After installation, open a terminal and type 'python --version' to verify it works. You can now create a new file called hello.py and write print('Hello World') to run your first program."} | |
| {"label": "wikipedia_style", "text": "The mitochondrion is a double-membrane-bound organelle found in most eukaryotic cells. It is best known as the site of cellular respiration, the process by which cells generate most of their energy in the form of adenosine triphosphate (ATP). Mitochondria are named from the Greek mitos (thread) and chondros (granule). They were first described by Richard Altmann in 1890."} | |
| {"label": "blog_post_med", "text": "I've been using Docker for about six months now and here are some things I wish I had known earlier. First, your Dockerfile order matters a lot \u2014 put your most frequently changing layers last. Second, multi-stage builds can dramatically reduce your final image size. Third, .dockerignore is not optional, it's essential. I've seen builds take 3x longer because someone forgot to exclude node_modules."} | |
| {"label": "textbook_explanation", "text": "Gradient descent is an optimization algorithm used to minimize a function by iteratively moving in the direction of steepest descent as defined by the negative of the gradient. In machine learning, we use gradient descent to update the parameters of our model. The learning rate determines the size of the steps taken toward the minimum. If the learning rate is too large, the algorithm may overshoot the minimum and diverge. If it is too small, convergence will be slow. Common variants include stochastic gradient descent (SGD), which uses a single random example per step, and adaptive methods like Adam, which maintain per-parameter learning rates based on the first and second moments of the gradients."} | |
| {"label": "detailed_lesson", "text": "A transformer is a neural network architecture that relies entirely on self-attention mechanisms to process sequences, without using recurrence or convolution. The key innovation is the scaled dot-product attention: given query Q, key K, and value V matrices, attention is computed as softmax(QK^T / sqrt(d_k)) * V. The division by the square root of the key dimension prevents the dot products from growing too large in magnitude, which would push the softmax into regions with extremely small gradients. Multi-head attention extends this by learning multiple sets of Q, K, V projections, allowing the model to attend to information from different representation subspaces at different positions. Each head operates on a reduced dimension d_k = d_model / n_heads, so the total computational cost is comparable to single-head attention with the full dimension."} | |
| {"label": "research_summary", "text": "The paper introduces a novel approach to language model compression using low-rank factorization of the attention weight matrices. By decomposing each attention head's query, key, and value projections into products of lower-rank matrices, the authors reduce parameter count by 40% while maintaining less than 1% perplexity degradation on the WikiText-103 benchmark. The method is particularly effective for smaller models (under 1B parameters), where the rank deficiency of attention matrices is more pronounced. The authors also demonstrate that the compression is compatible with knowledge distillation, enabling a two-stage pipeline that achieves 60% parameter reduction with only 2% perplexity increase. Code and model weights are released for reproducibility."} | |
| {"label": "comprehensive_guide", "text": "When training a language model from scratch, several hyperparameters interact in non-obvious ways. The learning rate schedule is critical: a warmup phase of 1-2% of total steps prevents early divergence, followed by cosine decay to near-zero. The batch size determines the effective learning rate \u2014 doubling the batch size typically requires a ~sqrt(2) increase in LR to maintain the same optimization dynamics. Context length affects both memory and quality: longer sequences provide more context for the model to learn from, but also increase the computational cost quadratically (or linearly with efficient attention). For models under 100M parameters, a context of 512-1024 tokens is usually sufficient; beyond that, the marginal benefit decreases. The vocabulary size is another key decision: byte-level tokenizers are simple and language-agnostic but produce longer sequences, while BPE tokenizers with 32k-50k vocabulary produce shorter sequences at the cost of a larger embedding table."} | |