infer-camembert

Python implementation for text classification inference with CamemBERT fine-tuned models

This is a simple Python implementation for the inference step of a fine-tuned text classification model based on Transformer's camembert-base model and saved in HuggingFace™.

Usage

$ pip install infer-camembert

For a private model, you must provide your HuggingFace token, either as an environment variable or under the ~/.huggingface folder:

$ HUGGINGFACE_TOKEN=<value> python3 -m infercamembert --input=example.jsonl --dictionary=labels.json --model="your-public-or-private-model-on-huggingface" --threshold=0.1 > results.jsonl

Inputs must be in the form of a dict with the keys being your unique IDs and the values the text on which to perform inference, eg.

{
  "id1": "Very nice time spent in a gorgeous site.",
  "id2": "Still a problem after three years: intolerable!!!!!!",
}

The same thing goes for the dictionary of labels where the keys should be your short custom labels and the value their corresponding long labels, eg.

{
  "label0": "undefined",
  "label1": "pleasure",
  "label2": "fun",
  "label3": "anger",
}

The results are presented as an array of predictions per input line, eg.

[
  {
    "id": "id1",
    "text": "Very nice time spent in a gorgeous site.", 
    "labels": [
      "pleasure",
      "fun"
    ]
  },
  {
    "id": "id2",
    "text": "Still a problem after three years: intolerable!!!!!!",
    "labels": [
      "anger"
    ]
  }
]

Used as a Python library:

from infercamembert import infer, Labels, ModelParameters

inputs = {
    "id1": "Very nice time spent in a gorgeous site.",
    "id2": "Still a problem after three years: intolerable!!!!!!",
}
labels = Labels(
    {
        "label0": "undefined",
        "label1": "pleasure",
        "label2": "fun",
        "label3": "anger",
    }
)
params = ModelParameters("your-public-or-private-model-on-huggingface", 0.1)
outputs = infer(inputs, labels, params)

Each call to infer loads the model unless you pass one loaded beforehand with load_model, which is the way to go when inferring in several rounds (the texts are then processed in chunks of batch_size, padded to the longest text of each chunk):

from infercamembert import infer, load_model, Labels, ModelParameters

params = ModelParameters("your-public-or-private-model-on-huggingface", 0.1)
loaded = load_model(params)
for inputs in rounds_of_inputs:
    outputs = infer(inputs, labels, params, loaded=loaded, batch_size=64)

License

This module is distributed under a MIT license.
See the LICENSE file.

---

© 2024-2026 Cyril Dever. All rights reserved.