Text Classifier

The text_classifier model assigns predefined labels to text inputs. It uses sequence classification models served via vLLM to score inputs against a fixed label set.

Dependencies

See Dependency Management for the complete provider matrix.

✦₊⁺ Overview

Text classification models score each input against a fixed set of labels and return the best match with a confidence score. They enable:

Sentiment Analysis: Positive, negative, neutral
Intent Detection: Categorize user intent without retraining
Topic Routing: Direct inputs to the right handler
Content Moderation: Detect toxic, spam, or unsafe content

1. Quick Start

Example

import msgflux as mf

# Labels defined at construction time
model = mf.Model.text_classifier(
    "vllm/jason9693/Qwen2.5-1.5B-apeach",
    labels=["toxic", "not_toxic"]
)

result = model("This product is absolutely amazing!").consume()
print(result)
# {"label": "not_toxic", "score": 0.97}

2. Supported Providers

vLLM (self-hosted)

vLLM serves SequenceClassification models via the /classify endpoint. Start the server with --task classify.

Example

Qwen2.5 ClassifierModernBERT Classifier

import msgflux as mf

# jason9693/Qwen2.5-1.5B-apeach — cited in vLLM docs as canonical classify example
# vllm serve jason9693/Qwen2.5-1.5B-apeach --task classify
model = mf.Model.text_classifier(
    "vllm/jason9693/Qwen2.5-1.5B-apeach",
    labels=["toxic", "not_toxic"],
    base_url="http://localhost:8000/v1"
)

import msgflux as mf

# Any fine-tuned ModernBERT sequence classifier works with vLLM (April 2025+)
# vllm serve davanstrien/ModernBERT-web-topics-1m --task classify
model = mf.Model.text_classifier(
    "vllm/davanstrien/ModernBERT-web-topics-1m",
    labels=["technology", "science", "sports", "politics"],
    base_url="http://localhost:8000/v1"
)

3. Batch Classification

Pass a list to classify multiple texts in a single call:

Example

import msgflux as mf

model = mf.Model.text_classifier(
    "vllm/jason9693/Qwen2.5-1.5B-apeach",
    labels=["urgent", "normal", "low_priority"],
    base_url="http://localhost:8000/v1"
)

tickets = [
    "System is completely down, customers cannot login",
    "Please update my billing address",
    "Would be nice to have dark mode someday"
]

results = model(tickets).consume()
# [
#   {"label": "urgent",       "score": 0.97},
#   {"label": "normal",       "score": 0.85},
#   {"label": "low_priority", "score": 0.91}
# ]

for ticket, result in zip(tickets, results):
    print(f"[{result['label']}] {ticket[:50]}...")

4. Async Support

Example

import msgflux as mf
import msgflux.nn.functional as F

model = mf.Model.text_classifier(
    "vllm/jason9693/Qwen2.5-1.5B-apeach",
    labels=["positive", "negative", "neutral"],
    base_url="http://localhost:8000/v1"
)

reviews = ["Great product!", "Terrible experience.", "It was okay."]

results = F.map_gather(
    model,
    args_list=[(r,) for r in reviews]
)

for review, result in zip(reviews, results):
    print(f"{review!r}: {result.consume()}")

5. Error Handling

Example

import msgflux as mf

model = mf.Model.text_classifier(
    "vllm/jason9693/Qwen2.5-1.5B-apeach",
    labels=["positive", "negative"],
    base_url="http://localhost:8000/v1"
)

try:
    result = model("Some text").consume()
    print(result["label"], result["score"])
except ImportError:
    print("Provider not installed")
except Exception as e:
    print(f"Classification failed: {e}")