NCP-GENL Data Preparation — All Questions With Answers

Question 1

When curating a high-quality instruction tuning dataset for fine-tuning a Large Language Model, which TWO factors are essential for maintaining model performance and safety?

Accepted Answer

Diversity of instruction formats to prevent overfitting to specific task structures.. High-quality data curation requires a balance between diversity of instruction types and strict adherence to safety guidelines. Ensuring the data is representative of real-world use cases prevents overfitting, while careful filtering of toxic or biased content ensures the model aligns with expected safety standards. These steps are foundational to successful fine-tuning in any professional NVIDIA-accelerated AI project.

Answer

Maximum possible volume of raw, uncleaned web-scraped data.

Answer

Removing all professional terminology to ensure the model remains accessible to general users.

Answer

Including non-English data to maximize the model's multilingual capabilities.

Question 2

In the context of NVIDIA NeMo, which THREE steps are foundational to the data preparation pipeline for pre-training large language models?

Accepted Answer

Normalization of unicode characters and formatting.. Data preparation in NeMo involves a robust pipeline consisting of filtering, deduplication, and tokenization. These steps ensure that the training data is clean, efficient, and formatted correctly for the underlying GPU architecture. Adhering to these standard steps minimizes computational waste and improves the final quality of the pre-trained model by focusing on high-quality text representations.

Answer

Implementing a custom GPU-based real-time sentiment analysis layer.

Answer

Replacing all punctuation marks with unique tokens.

Question 3

When preparing datasets for Retrieval-Augmented Generation (RAG), which THREE factors are essential to ensure efficient retrieval on an NVIDIA-accelerated vector database?

Accepted Answer

Consistent metadata tagging for efficient filtering and partitioning.. Efficient RAG depends on the synergy between the indexing strategy and the query structure. Using appropriate embedding models ensures semantic relevance, while metadata filtering narrows the search space effectively. Finally, optimizing chunk sizes maintains a balance between retrieval speed and the richness of the retrieved context, which are all vital for maintaining low latency and high quality in production RAG systems.

Answer

Encoding all data into 1-bit quantized vectors to save space.

Answer

Indexing only the top 10% of the most popular documents.

Question 4

You are curating a pretraining corpus with NVIDIA NeMo Curator and need to remove low-quality documents before tokenization. Which two NeMo Curator heuristic filtering criteria are appropriate for this goal? (Choose two.)

Accepted Answer

Filtering documents whose punctuation ratio or symbol-to-word ratio exceeds a threshold.. NeMo Curator's heuristic quality filters include punctuation and symbol-to-word ratios, which catch boilerplate and corrupted text, and stop-word ratios, which catch keyword lists and unnatural text. Both target content that degrades pretraining without discarding legitimate long-form or multilingual documents. Applying these two filters before tokenization removes noise while preserving valuable text, directly meeting the scenario's requirement to clean the corpus.

Answer

Filtering documents that contain any named entity recognized by an NER model.

Answer

Filtering documents that contain more than one language.

Answer

Filtering documents that exceed a maximum token count of 512 tokens.

Question 5

When preparing a proprietary technical manual dataset for a RAG pipeline, which data preprocessing step is most critical to ensure the LLM avoids hallucinations regarding specific product configurations?

Accepted Answer

Implementing document-aware recursive character splitting with overlapping segments and metadata tagging.. Chunking strategies and metadata tagging ensure that context retrieval is precise. By preserving technical hierarchy and associating data with specific product versions, the LLM retrieves ground-truth documentation rather than generic information. This reduces hallucinations by constraining the search space to relevant, version-controlled text blocks, directly impacting the accuracy and reliability of downstream inference tasks in enterprise NVIDIA-based AI deployments.

Answer

Converting all text to lowercase to ensure uniformity in vector embeddings.

Answer

Performing aggressive stop-word removal to reduce the dimensionality of the vector space.

Answer

Applying basic sentence tokenization based strictly on periods to create uniform chunks.

Question 6

Refer to the exhibit. A team is preparing log data for a RAG-based troubleshooting assistant. Given the configuration, what is the most significant risk during the retrieval phase?

Accepted Answer

Setting the overlap to 0 will break contextual continuity between consecutive log entries.. The lack of overlap between chunks is critical. In log data, sequential events are often related; by having zero overlap, the system loses the transition between log lines that might contain a causal link. This fragmentation forces the model to interpret isolated snapshots, hindering its ability to reconstruct the sequence of errors or system states, which is vital for effective root-cause analysis.

Answer

The use of 'nv-embed-v1' will cause OOM errors during the indexing phase.

Answer

The chunk size of 4096 tokens is too small for modern log analysis.

Answer

The 'regex_mask_all' policy will cause the embedding model to fail during vectorization.

Question 7

Which technique is most effective for mitigating data leakage during the training of an LLM on time-series-related document data?

Accepted Answer

Chronological splitting of the dataset based on a cutoff date.. Preventing data leakage in temporal datasets requires strict chronological separation between training, validation, and testing sets. If future information is inadvertently included in the training data, the model will 'cheat' by observing future events during training, leading to artificially inflated performance metrics that do not generalize to actual production deployment scenarios where future data is unavailable.

Answer

Random shuffling of all documents regardless of their timestamps.

Answer

Oversampling the minority class in the training set to improve balance.

Answer

Applying aggressive data normalization to all numerical values.

Question 8

Refer to the exhibit. What is the intended outcome of this data cleaning configuration for a Large Language Model pre-training corpus?

Accepted Answer

It optimizes for the removal of low-quality or nonsensical text while minimizing redundancy.. This configuration aims to remove low-quality text that fails to meet minimum length requirements or exhibits high perplexity (indicating gibberish or low coherence). Simultaneously, MinHash deduplication identifies and removes near-duplicate documents exceeding a 95% similarity threshold. This cleanup process is vital for pre-training, as it filters out low-value, noisy data that could impede model convergence and general quality during the training cycle.

Answer

It specifically targets the removal of personally identifiable information (PII).

Answer

It enforces a strict length-based chunking strategy for all documents.

Answer

It converts all text to a vector space representation before filtering.

Question 9

Why is 'tokenization stability' a critical metric when preparing data for NVIDIA-based LLM deployment?

Accepted Answer

It prevents unexpected input shifts between training and inference environments.. Tokenization stability ensures that the same input text consistently maps to the same sequence of tokens across different environments or library versions. If tokenization is inconsistent, the model might receive unexpected inputs compared to what it observed during training, leading to severe performance degradation. For NVIDIA-based deployments, deterministic tokenization is essential for maintaining production-level reliability and predictable model behavior across various inference pipelines.

Answer

It guarantees that the model will always generate the same output for a given prompt.

Answer

It ensures that the GPU memory usage remains constant during the training process.

Answer

It reduces the total number of parameters required for the embedding layer.

Question 10

When preparing unstructured documentation for a high-performance retrieval system, which approach best balances index size and retrieval relevance?

Accepted Answer

Implementing sliding window chunking with semantic overlap based on document structure.. Utilizing a sliding window approach with semantic overlap ensures that retrieved chunks maintain context. By carefully selecting chunk size and overlap, engineers can optimize the index size to avoid redundant storage while ensuring that the semantic units of the text are not cut off. This balance is vital for maximizing the accuracy of RAG systems running on NVIDIA infrastructure, where memory efficiency is paramount.

Answer

Using extremely large chunks to ensure that every document is a single vector.

Answer

Storing every sentence as an individual chunk in the vector database.

Answer

Removing all overlaps to keep the index size at the absolute minimum possible.

Question 11

When fine-tuning an LLM to follow specific safety protocols, why is the inclusion of 'adversarial' examples in the training data considered a best practice?

Accepted Answer

To ensure the model learns to refuse requests that violate established safety policies.. Adversarial examples test the model's ability to maintain safety boundaries even when prompted with manipulative or malicious inputs. By training on these examples, the model learns to identify and refuse requests that violate safety protocols. This proactive preparation is essential for deploying LLMs in enterprise environments where maintaining strict safety and compliance standards is required, protecting against sophisticated prompt injection and bypass techniques.

Answer

To artificially inflate the model's perplexity scores during evaluation.

Answer

To increase the model's fluency in generating complex technical instructions.

Answer

To allow the model to learn and reproduce the adversarial techniques during generation.

Question 12

When assessing the quality of a dataset for instruction fine-tuning, which TWO metrics or methods are considered most reliable for measuring dataset diversity?

Accepted Answer

Embedding-based clustering to visualize topical coverage across the corpus.. Dataset diversity ensures that the model encounters a broad range of topics and linguistic patterns, which is critical for generalization. Using embedding-based clustering allows for the identification of thematic coverage, while perplexity distribution analysis helps assess whether the dataset contains a balance of common and complex structures. These methods together provide a quantitative view of the data's breadth, reducing the risk of bias or overfitting.

Answer

Calculating the total number of words in the dataset.

Answer

Measuring the average response length for every instruction.

Answer

Checking if the dataset is solely in ASCII format.

Question 13

What is the primary function of data 'normalization' in the context of preparing inputs for a Transformer model?

Accepted Answer

Standardizing the text to ensure consistent interpretation of tokens.. Normalization, such as standardizing whitespace, handling special tokens, and ensuring consistent character encodings, ensures that the model interprets input text in a predictable manner. By removing noise that doesn't contribute to semantic meaning, the model can focus its capacity on learning complex linguistic patterns. This is a standard and essential step in any high-performance AI data pipeline using NVIDIA accelerated computing.

Answer

Increasing the complexity of the input text to help the model learn more features.

Answer

Encrypting the dataset to protect sensitive information during training.

Answer

Compressing the dataset to reduce storage space on the GPU disk.

Question 14

Why is it important to perform 'domain-specific' data cleaning when preparing a corpus for fine-tuning a medical LLM?

Accepted Answer

Generic cleaning often removes or modifies critical domain-specific nomenclature.. Medical text contains specific jargon, abbreviations, and relationships that generic cleaning might misinterpret. Generic tools often remove or alter terms that are critical to medical context, such as drug names or procedural codes. By using domain-specific cleaning, you ensure that the model retains the precise terminology necessary for high-stakes, accurate clinical reasoning in medical applications.

Answer

It helps to anonymize the patient data by replacing all medical terms with generic labels.

Answer

It ensures that the dataset size is significantly reduced to fit in GPU cache.

Answer

It automatically corrects all scientific inaccuracies present in the original documents.

Question 15

Refer to the exhibit. What is the impact of this filter on the training corpus?

Accepted Answer

It discards documents that are overly repetitive or lack linguistic richness.. This filter removes documents with low lexical diversity, which often contain repetitive, low-value, or 'boilerplate' text. Such content provides little signal for the model to learn meaningful language patterns. By enforcing a minimum diversity threshold, you ensure the corpus consists of richer, more informative language, which typically leads to better convergence and higher quality output in the trained model.

Answer

It limits the vocabulary size to 20% of the original content.

Answer

It forces the model to use 20% more computation for tokenization.

Answer

It reduces the training corpus size by exactly 20%.

Question 16

When training a model for a highly technical domain with a scarcity of high-quality data, which data augmentation strategy is most likely to preserve the model's reliability?

Accepted Answer

Generating synthetic examples based on ground-truth technical documentation templates.. In technical domains, synthetic data generation must be grounded in existing, verified documents to avoid 'hallucinating' technical facts. Using LLMs to paraphrase or summarize existing high-quality technical content while maintaining strict constraints ensures that the new data follows the same logic and terminology. This method expands the training set while minimizing the risk of introducing incorrect facts, which is essential for high-stakes technical domains.

Answer

Randomly masking 50% of the words in the training documents.

Answer

Using a smaller, unverified LLM to generate completely new technical scenarios.

Answer

Translating the dataset into multiple languages using a generic online translator.

Question 17

You are fine-tuning a LLM on a large corpus of technical documentation. You notice the model struggles with domain-specific terminology that frequently appeared in the training data but was inconsistently formatted. Which data preparation technique best addresses this issue?

Accepted Answer

Applying consistent text normalization and domain-specific tokenization rules.. Consistent normalization, such as lemmatization or standardizing case and special characters, ensures the tokenizer treats synonymous terms identically. In NVIDIA NeMo workflows, data quality directly impacts convergence speed and model accuracy. By reducing vocabulary noise during the preprocessing stage, the model can dedicate more capacity to learning semantic relationships rather than mapping variants of the same technical term to different embedding spaces, ultimately improving downstream performance on domain-specific benchmarks.

Answer

Increasing the batch size during the training phase.

Answer

Expanding the model's depth with additional transformer layers.

Answer

Reducing the learning rate to prevent overfitting on the noise.

Question 18

You are preparing a massive dataset for training a NeMo-based LLM. Which TWO data preprocessing steps are critical to prevent data leakage and ensure model quality?

Accepted Answer

Performing temporal or content-based splitting to isolate the validation set.. Data leakage occurs when test data is inadvertently included in the training set, leading to inflated performance metrics. Deduplication is equally vital, as repetitive data causes the model to memorize samples rather than generalize. In NVIDIA workflows, these steps are typically performed via distributed scripts on the cluster before tokenization, ensuring that the model learns unique, non-overlapping information across all shards.

Answer

Increasing the frequency of data augmentation in the training pipeline.

Answer

Adding metadata tags to all training sequences for classification.

Answer

Converting all text to lowercase to increase vocabulary efficiency.

Question 19

Refer to the exhibit. You are reviewing the configuration file for a data preprocessing pipeline. Why is the 'min_words' filter set to 50 in the context of LLM training?

Accepted Answer

It removes low-quality snippets to ensure meaningful semantic context.. The 'min_words' filter removes low-information or malformed snippets that lack sufficient context for effective transformer learning. In large-scale training, such as those performed on NVIDIA H100 GPU clusters, including very short strings increases noise and consumes valuable compute cycles without contributing to meaningful semantic representation. Setting a minimum length ensures the model learns from coherent passages rather than disjointed fragments, promoting better structural understanding of the training corpus.

Answer

It limits the memory footprint of the dataloader.

Answer

It enables faster tokenization by reducing the number of input files.

Answer

It forces the model to ignore PII-heavy documents.

Question 20

An enterprise is preparing a massive multi-terabyte corpus of specialized PDF documents containing technical schematics and dense tables for training a domain-specific LLM using NeMo Curator. During the document extraction pipeline, you notice that tabular data layouts are scrambled into unstructured string tokens, destroying spatial relationships. Which data preparation strategy should you implement first within NeMo Curator to preserve tabular integrity before tokenization?

Accepted Answer

Deploy NeMo Curator's layout-aware PDF extraction modules featuring visual bounding box detection to parse tables into structured Markdown formatting.. NeMo Curator provides specialized PDF extraction utilities that leverage advanced computer vision and layout parsing models to accurately identify bounding boxes for tables and figures. Preserving tabular markdown structures prevents spatial collapse, ensuring downstream tokenizers capture relational data accurately. This step is critical in domain-specific LLM training because scrambled tables introduce severe noise that degrades reasoning capabilities.

Answer

Apply aggressive regex-based whitespace removal across all extracted string buffers to normalize sentence boundaries prior to embedding generation.

Answer

Increase the chunk size parameter in the tokenizer configuration so that entire table blocks fit inside a single oversized context window.

Answer

Convert all PDF pages into low-resolution JPEG images to bypass text extraction errors and feed raw pixels directly into a text-only causal language model.

NCP-GENL Data Preparation — All Questions With Answers

When curating a high-quality instruction tuning dataset for fine-tuning a Large Language Model, which TWO factors are essential for maintaining model performance and safety?

In the context of NVIDIA NeMo, which THREE steps are foundational to the data preparation pipeline for pre-training large language models?

When preparing datasets for Retrieval-Augmented Generation (RAG), which THREE factors are essential to ensure efficient retrieval on an NVIDIA-accelerated vector database?

You are curating a pretraining corpus with NVIDIA NeMo Curator and need to remove low-quality documents before tokenization. Which two NeMo Curator heuristic filtering criteria are appropriate for this goal? (Choose two.)

When preparing a proprietary technical manual dataset for a RAG pipeline, which data preprocessing step is most critical to ensure the LLM avoids hallucinations regarding specific product configurations?

Refer to the exhibit. A team is preparing log data for a RAG-based troubleshooting assistant. Given the configuration, what is the most significant risk during the retrieval phase?

Exhibit

Which technique is most effective for mitigating data leakage during the training of an LLM on time-series-related document data?

Refer to the exhibit. What is the intended outcome of this data cleaning configuration for a Large Language Model pre-training corpus?

Exhibit

Why is 'tokenization stability' a critical metric when preparing data for NVIDIA-based LLM deployment?

When preparing unstructured documentation for a high-performance retrieval system, which approach best balances index size and retrieval relevance?

When fine-tuning an LLM to follow specific safety protocols, why is the inclusion of 'adversarial' examples in the training data considered a best practice?

When assessing the quality of a dataset for instruction fine-tuning, which TWO metrics or methods are considered most reliable for measuring dataset diversity?

What is the primary function of data 'normalization' in the context of preparing inputs for a Transformer model?

Why is it important to perform 'domain-specific' data cleaning when preparing a corpus for fine-tuning a medical LLM?

Refer to the exhibit. What is the impact of this filter on the training corpus?

Exhibit

When training a model for a highly technical domain with a scarcity of high-quality data, which data augmentation strategy is most likely to preserve the model's reliability?

You are fine-tuning a LLM on a large corpus of technical documentation. You notice the model struggles with domain-specific terminology that frequently appeared in the training data but was inconsistently formatted. Which data preparation technique best addresses this issue?

You are preparing a massive dataset for training a NeMo-based LLM. Which TWO data preprocessing steps are critical to prevent data leakage and ensure model quality?

Refer to the exhibit. You are reviewing the configuration file for a data preprocessing pipeline. Why is the 'min_words' filter set to 50 in the context of LLM training?

Exhibit

You are curating a 2 TB corpus of NVIDIA technical documentation and Python code for continued pretraining of a NeMo-based LLM. A colleague proposes filtering out any document containing the token sequence 'CUDA' to reduce hardware-specific bias. What is the most appropriate response?

You are using NVIDIA NeMo Curator to filter a 600 GB web-crawl corpus before pre-training. Your team wants to remove exact duplicates and near-duplicates to reduce memorization and speed up training. Which NeMo Curator stage should you apply?

You are preparing a large instruction-tuning dataset with NVIDIA NeMo Curator. The dataset contains many near-duplicate instruction-response pairs that differ only in punctuation or minor wording. Which NeMo Curator stage should you apply to remove these near-duplicates before fine-tuning?

A team is preparing a customer-support chat dataset for instruction fine-tuning of an LLM. The raw data contains HTML tags, inconsistent date formats, and emoji. Which data preparation step should be performed first to make the text usable for tokenization?

A team is building a NeMo-based LLM pipeline and must tokenize a corpus that mixes English, Japanese, and Python source code. They plan to train a custom tokenizer with NVIDIA NeMo. Which tokenizer configuration best supports all three content types without excessive sequence length?

You are curating instruction-tuning data for an NVIDIA NIM-deployed LLM. The raw dataset contains many near-duplicate instruction-response pairs that differ only by punctuation and whitespace. Which data preparation step is most appropriate to remove these before fine-tuning?

You are preparing a large instruction-tuning dataset for an LLM using NVIDIA NeMo. You need to ensure the dataset supports efficient training and evaluation. Which TWO steps are essential in the data preparation pipeline? (Choose two.)

You are preparing a multilingual corpus for pretraining with NVIDIA NeMo. The dataset contains documents in 40 languages, but the tokenizer was trained primarily on English. Which data preparation action best ensures that non-English text is represented efficiently during tokenization?

You are preparing a multilingual corpus for pre-training an LLM. The corpus contains documents in English, Spanish, and German, but the English portion is 80% of the data. You want to avoid the model becoming biased toward English. Which data preparation technique should you apply?

You are building a pretraining dataset from a large collection of source-code repositories for an NVIDIA NeMo LLM. The data includes many files with licenses, generated code, and minified JavaScript. Which NeMo Curator-based approach best improves code data quality before tokenization?

You are preparing a customer-support dataset for fine-tuning an LLM with NVIDIA NeMo. The raw data includes personally identifiable information such as names, email addresses, and phone numbers. Which data preparation step must be performed before training to comply with privacy requirements?

You are preparing a dataset for supervised fine-tuning (SFT) of an NVIDIA NeMo LLM to follow instructions. Which TWO data preparation practices are essential to ensure the model learns to generalize rather than memorize? (Choose two.)

You are preparing a dataset of support tickets for a RAG system using NVIDIA NeMo. Many tickets are short and contain little context, which hurts retrieval quality. Which data preparation technique best improves retrieval by enriching each ticket with related information before embedding?

NCP-GENL Data Preparation — All Questions With Answers

When curating a high-quality instruction tuning dataset for fine-tuning a Large Language Model, which TWO factors are essential for maintaining model performance and safety?

In the context of NVIDIA NeMo, which THREE steps are foundational to the data preparation pipeline for pre-training large language models?

When preparing datasets for Retrieval-Augmented Generation (RAG), which THREE factors are essential to ensure efficient retrieval on an NVIDIA-accelerated vector database?

You are curating a pretraining corpus with NVIDIA NeMo Curator and need to remove low-quality documents before tokenization. Which two NeMo Curator heuristic filtering criteria are appropriate for this goal? (Choose two.)

When preparing a proprietary technical manual dataset for a RAG pipeline, which data preprocessing step is most critical to ensure the LLM avoids hallucinations regarding specific product configurations?

Refer to the exhibit. A team is preparing log data for a RAG-based troubleshooting assistant. Given the configuration, what is the most significant risk during the retrieval phase?

Exhibit

Which technique is most effective for mitigating data leakage during the training of an LLM on time-series-related document data?

Refer to the exhibit. What is the intended outcome of this data cleaning configuration for a Large Language Model pre-training corpus?

Exhibit

Why is 'tokenization stability' a critical metric when preparing data for NVIDIA-based LLM deployment?

When preparing unstructured documentation for a high-performance retrieval system, which approach best balances index size and retrieval relevance?

When fine-tuning an LLM to follow specific safety protocols, why is the inclusion of 'adversarial' examples in the training data considered a best practice?

When assessing the quality of a dataset for instruction fine-tuning, which TWO metrics or methods are considered most reliable for measuring dataset diversity?

What is the primary function of data 'normalization' in the context of preparing inputs for a Transformer model?

Why is it important to perform 'domain-specific' data cleaning when preparing a corpus for fine-tuning a medical LLM?

Refer to the exhibit. What is the impact of this filter on the training corpus?

Exhibit

When training a model for a highly technical domain with a scarcity of high-quality data, which data augmentation strategy is most likely to preserve the model's reliability?

You are fine-tuning a LLM on a large corpus of technical documentation. You notice the model struggles with domain-specific terminology that frequently appeared in the training data but was inconsistently formatted. Which data preparation technique best addresses this issue?

You are preparing a massive dataset for training a NeMo-based LLM. Which TWO data preprocessing steps are critical to prevent data leakage and ensure model quality?

Refer to the exhibit. You are reviewing the configuration file for a data preprocessing pipeline. Why is the 'min_words' filter set to 50 in the context of LLM training?

Exhibit

You are curating a 2 TB corpus of NVIDIA technical documentation and Python code for continued pretraining of a NeMo-based LLM. A colleague proposes filtering out any document containing the token sequence 'CUDA' to reduce hardware-specific bias. What is the most appropriate response?

You are using NVIDIA NeMo Curator to filter a 600 GB web-crawl corpus before pre-training. Your team wants to remove exact duplicates and near-duplicates to reduce memorization and speed up training. Which NeMo Curator stage should you apply?

You are preparing a large instruction-tuning dataset with NVIDIA NeMo Curator. The dataset contains many near-duplicate instruction-response pairs that differ only in punctuation or minor wording. Which NeMo Curator stage should you apply to remove these near-duplicates before fine-tuning?

A team is preparing a customer-support chat dataset for instruction fine-tuning of an LLM. The raw data contains HTML tags, inconsistent date formats, and emoji. Which data preparation step should be performed first to make the text usable for tokenization?

A team is building a NeMo-based LLM pipeline and must tokenize a corpus that mixes English, Japanese, and Python source code. They plan to train a custom tokenizer with NVIDIA NeMo. Which tokenizer configuration best supports all three content types without excessive sequence length?

You are curating instruction-tuning data for an NVIDIA NIM-deployed LLM. The raw dataset contains many near-duplicate instruction-response pairs that differ only by punctuation and whitespace. Which data preparation step is most appropriate to remove these before fine-tuning?

You are preparing a large instruction-tuning dataset for an LLM using NVIDIA NeMo. You need to ensure the dataset supports efficient training and evaluation. Which TWO steps are essential in the data preparation pipeline? (Choose two.)

You are preparing a multilingual corpus for pretraining with NVIDIA NeMo. The dataset contains documents in 40 languages, but the tokenizer was trained primarily on English. Which data preparation action best ensures that non-English text is represented efficiently during tokenization?

You are preparing a multilingual corpus for pre-training an LLM. The corpus contains documents in English, Spanish, and German, but the English portion is 80% of the data. You want to avoid the model becoming biased toward English. Which data preparation technique should you apply?

You are building a pretraining dataset from a large collection of source-code repositories for an NVIDIA NeMo LLM. The data includes many files with licenses, generated code, and minified JavaScript. Which NeMo Curator-based approach best improves code data quality before tokenization?

You are preparing a customer-support dataset for fine-tuning an LLM with NVIDIA NeMo. The raw data includes personally identifiable information such as names, email addresses, and phone numbers. Which data preparation step must be performed before training to comply with privacy requirements?

You are preparing a dataset for supervised fine-tuning (SFT) of an NVIDIA NeMo LLM to follow instructions. Which TWO data preparation practices are essential to ensure the model learns to generalize rather than memorize? (Choose two.)

You are preparing a dataset of support tickets for a RAG system using NVIDIA NeMo. Many tickets are short and contain little context, which hurts retrieval quality. Which data preparation technique best improves retrieval by enriching each ticket with related information before embedding?