huggingface

Multi-Task NLP via Hugging Face: Model Selection by Task Type

Different NLP tasks need different models with different output formats. Summarization uses BART and returns summary_text. Text generation uses Mixtral and returns generated_text. Sentiment analysis uses DistilBERT and returns label. This workflow routes to the right model based on the task type, calls the Hugging Face Inference API, and parses the task-specific output format.

Workflow

text, task
   │
   ▼
┌──────────────────┐
│ hf_select_model  │  Map task type to model ID + params
└────────┬─────────┘
         │  modelId, formattedInput, parameters, options
         ▼
┌──────────────────┐
│ hf_inference     │  POST to api-inference.huggingface.co
└────────┬─────────┘
         │  rawOutput
         ▼
┌──────────────────┐
│ hf_format_result │  Extract task-specific field
└──────────────────┘
         │
         ▼
   result, model, task

Workers

HfSelectModelWorker (hf_select_model) -- Maps task types to model IDs via a static TASK_MODEL_MAP: "text-generation" -> "mistralai/Mixtral-8x7B-Instruct-v0.1", "summarization" -> "facebook/bart-large-cnn", "sentiment-analysis" -> "distilbert-base-uncased-finetuned-sst-2-english", "translation" -> "Helsinki-NLP/opus-mt-en-fr". Unknown task types default to text-generation. Returns inference parameters: max_new_tokens: 250, temperature: 0.6, top_p: 0.9, repetition_penalty: 1.1, do_sample: true. Also sets options: {wait_for_model: true, use_cache: false}.

HfInferenceWorker (hf_inference) -- Checks HUGGINGFACE_TOKEN at construction. In live mode, POSTs {"inputs": text} to https://api-inference.huggingface.co/models/<modelId> with a 10-second connect timeout and 120-second request timeout, Authorization: Bearer <token>. In fallback mode, returns a hardcoded summarization result about a hybrid work study (2,500 workers, 12 countries, 18%/23% productivity improvements).

HfFormatResultWorker (hf_format_result) -- Extracts the result based on task type: "summarization" reads output.get(0).get("summary_text"), "sentiment-analysis" reads .get("label"), everything else reads .get("generated_text") with a .toString() fallback. Logs the first 60 characters of the result.

Tests

17 tests cover model selection for all task types, inference in both modes, and format extraction for each output shape.

Further Reading

• RUNNING.md -- how to build and run this example