---

name: nlp-and-generation description: "Use Libra for text classification, summarization, named entity recognition, GPT-2 text generation, image captioning, and NLTK/TextBlob runtime setup." disable-model-invocation: true metadata: disco-role: operating license: MIT

NLP and Generation with Libra

Load this sub-skill when the task uses Libra's text and language-generation APIs: sentiment/text classification, summarization, GPT-2 text generation, named entity recognition, image captioning, vocabulary inspection, or NLTK/TextBlob runtime setup.

What this sub-skill owns

• text_classification_query(...) and classify_text(...)
• summarization_query(...) and get_summary(...)
• generate_text(...)
• named_entity_query(...)
• image_caption_query(...) and generate_caption(...)
• NLTK corpora, TextBlob/POS-tagging assumptions, HuggingFace/TensorFlow model downloads, and NLP preprocessing helper behavior

Trigger phrases

Use this route when a user asks to:

• train a sentiment or text classifier from a CSV and classify new text
• fine-tune or use summarization on text/summary pairs
• generate text from a file or prefix prompt
• detect named entities in a text column
• caption images from a CSV of image paths and captions
• fix missing NLTK corpora, TextBlob tagging, or HuggingFace model download failures

Bundled references

• references/api-reference.md for methods, model keys, and parameters
• references/workflows.md for end-to-end recipes
• references/data-formats.md for text, summary, generation, NER, and captioning data shapes
• references/runtime-setup.md for NLTK, TextBlob, transformers, and optional GPU setup
• references/troubleshooting.md for common NLP/generation failures

Bundled scripts

• scripts/prepare_nltk_corpora.py checks required corpora and can optionally download them when network access is explicitly allowed.
• scripts/smoke_text_generation.py checks the generate_text API surface without forcing a GPT-2 download by default.

Operating notes

1. client.__init__ downloads NLTK punkt, averaged_perceptron_tagger, and stopwords; patch it only for non-NLP smoke checks that do not need corpora.
2. text_classification_query defaults to label column label; summarization_query defaults to label column summary.
3. generate_text(file_data=True) reads the client dataset path as a text file. Use file_data=False, prefix="..." when the user wants prompt-only generation.
4. Transformer workflows load HuggingFace models and can require network/cache access. State this before running them.
5. Image captioning is owned here because the public API lives with the NLP queries, but image-path/layout debugging should also consult sub-skills/vision-and-generative.

Cross-links

• Use the root skill for global install and pandas compatibility shims.
• Route tabular model selection and analyze()/plot workflows to sub-skills/tabular-modeling unless the model key is NLP-specific.
• Route image classification, CNN export, GANs, and image dataset layout checks to sub-skills/vision-and-generative.