Classify text

Frequently asked questions

Which models are the most accurate for text classification?

If you want the strongest overall accuracy and reasoning, use openai/gpt-5. It handles sentiment, topic, intent, moderation, and multi-step classification with high reliability.
If you want a second high-end option that excels at logic-heavy classification, use xai/grok-4. It is very good at subtle or ambiguous cases that require deeper context understanding.
For safety or policy driven classification, lucataco/gpt-oss-safeguard-20b is a strong alternative.

Which models offer the best balance of cost and performance?

For general purpose text classification at a reasonable cost, georgedavila/bart-large-mnli-classifier performs well because it uses zero shot NLI techniques.
If you want a lightweight option for sentiment tasks, curt-park/sentiment-analysis is inexpensive and very fast.

What should I use for sentiment analysis or brand monitoring?

For simple positive or negative sentiment tasks, curt-park/sentiment-analysis is the most direct model. It is fast, predictable, and well suited for large batches.
If you want sentiment plus explanation or nuanced emotional tone, openai/gpt-5 does well with few shot prompts.
If you want to mix sentiment with toxicity filtering, fofr/prompt-classifier can classify text for harmful content scores.

What should I use for topic classification or routing tasks?

For topic classification with flexible labels, georgedavila/bart-large-mnli-classifier is ideal because you can provide your own category list and it will rank them.
For topic routing where context is complex, xai/grok-4 works well because it understands long passages and subtle distinctions.
For policy governed topic routing, lucataco/gpt-oss-safeguard-20b lets you supply a written policy and returns labels according to your rules.

What should I use for classifying harmful or unsafe text?

For safety classification using custom rules, lucataco/gpt-oss-safeguard-20b is designed for policy driven safety labeling and returns structured rationales.
For toxicity scoring on a scale from zero to ten, fofr/prompt-classifier is a fine tuned Llama model built for identifying harmful or unsafe prompts.
For detecting safety violations using general reasoning, openai/gpt-5 can classify harmful or disallowed content with detailed explanations.

What is the difference between small fine tuned classifiers and large LLMs for text classification?

Small fine tuned models like curt-park/sentiment-analysis, georgedavila/bart-large-mnli-classifier, and fofr/prompt-classifier run quickly and cheaply but follow narrow tasks. They are ideal when you know your labels and want predictable output.
Large LLMs like openai/gpt-5 and xai/grok-4 can classify text using natural language instructions, support open ended labels, and explain their reasoning.
Policy based classification models like lucataco/gpt-oss-safeguard-20b sit between these extremes by offering structured safety classification without relying on closed systems.

What kinds of outputs do these models produce?

Small classification models such as curt-park/sentiment-analysis and georgedavila/bart-large-mnli-classifier usually return a label and sometimes a confidence score.
Moderation models such as fofr/prompt-classifier return a numeric toxicity score.
Large LLMs such as openai/gpt-5 or xai/grok-4 can return labels, explanations, or structured JSON depending on your prompt.
Policy based models such as lucataco/gpt-oss-safeguard-20b output both a label and a chain of reasoning.

How can I self host or publish my own text classifier on Replicate?

You can self host or publish your own classifier by building it with Cog and including a replicate.yaml file that defines inputs, outputs, prediction logic, and environment.
Models such as curt-park/sentiment-analysis and lucataco/gpt-oss-safeguard-20b demonstrate simple and complex setups.
Once you push the repository to Replicate, it runs on managed GPUs without extra configuration.

Can I use these models for commercial work?

Yes, as long as the license on each model page allows commercial use.
Open source models such as curt-park/sentiment-analysis, georgedavila/bart-large-mnli-classifier, and lucataco/gpt-oss-safeguard-20b can typically be used commercially.
Closed models such as openai/gpt-5 and xai/grok-4 have their own usage policies that you should review.

How do I run these models?

Upload or paste text into the model interface and choose any optional parameters.
For simple tasks like polarity detection, use curt-park/sentiment-analysis.
For flexible label sets, use georgedavila/bart-large-mnli-classifier.
For safety scoring, use fofr/prompt-classifier or lucataco/gpt-oss-safeguard-20b.
For complex or multi step classification, use openai/gpt-5 or xai/grok-4.

What should I know before running a job in this category?

Large LLMs such as openai/gpt-5 and xai/grok-4 cost more and take longer but give better reasoning.
Small models such as curt-park/sentiment-analysis and georgedavila/bart-large-mnli-classifier are faster and cheaper but have limited flexibility.
Safety models such as lucataco/gpt-oss-safeguard-20b can classify against your own custom policies.
Always check the model input format and consider testing a small batch first.

Any other tips for choosing between these models?

Use small specialized models such as curt-park/sentiment-analysis for narrow tasks where speed matters.
Use NLI based models such as georgedavila/bart-large-mnli-classifier for flexible zero shot label lists.
Use domain specific classifiers such as fofr/prompt-classifier for toxicity and prompt safety.
Use reasoning heavy LLMs such as openai/gpt-5 or xai/grok-4 when labels are ambiguous or when you need explanations.
Use policy driven safety models such as lucataco/gpt-oss-safeguard-20b when you must classify text according to internal rules or compliance requirements.