|
--- |
|
license: apache-2.0 |
|
language: |
|
- en |
|
base_model: |
|
- microsoft/deberta-v3-base |
|
--- |
|
# Slop Classifier for Roleplay Characters |
|
|
|
> This model can detect characters that are created using AI. |
|
|
|
Part of [CharGen](https://huggingface.co/kubernetes-bad/chargen-v2) project - it is used to detect and filter out low-effort, LLM-made characters intended for role playing. |
|
|
|
*Slop* refers to over-used phrases that models like GPT3.5 like to use very much and that do not add any value to the text. "Shivers down her spine", "enigma wrapped in mystery", "half-lidded eyes", etc. Classifier is trained on set of synthetic characters generated with GPT3.5 and GPT4, and a subset of CharGen dataset. |
|
|
|
## Usage |
|
|
|
```py |
|
from transformers import AutoTokenizer, AutoModelForSequenceClassification |
|
import torch |
|
from litserve import LitAPI, LitServer |
|
|
|
MODEL_NAME = "kubernetes-bad/character-slop-classifier" |
|
|
|
class CHARLitAPI(LitAPI): |
|
def setup(self, device): |
|
self.tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME) |
|
self.model = AutoModelForSequenceClassification.from_pretrained(MODEL_NAME) |
|
self.model.to(device) |
|
self.model.eval() |
|
|
|
def decode_request(self, request): |
|
if "text" in request: |
|
inputs = self.tokenizer(request["text"], return_tensors="pt", padding=True, truncation=True, max_length=512) |
|
elif "texts" in request: |
|
inputs = self.tokenizer(request["texts"], return_tensors="pt", padding=True, truncation=True, max_length=512) |
|
else: |
|
raise ValueError("Invalid request format. Expected 'text' or 'texts' field.") |
|
return inputs |
|
|
|
def predict(self, inputs): |
|
with torch.no_grad(): |
|
inputs = {k: v.to(self.model.device) for k, v in inputs.items()} |
|
outputs = self.model(**inputs) |
|
return outputs.logits |
|
|
|
def encode_response(self, logits): |
|
probabilities = torch.nn.functional.softmax(logits, dim=-1) |
|
if probabilities.shape[0] == 1: |
|
response = { |
|
"positive": probabilities[:, 1].item(), |
|
"negative": probabilities[:, 0].item() |
|
} |
|
else: |
|
response = [ |
|
{ |
|
"positive": prob[1].item(), |
|
"negative": prob[0].item() |
|
} |
|
for prob in probabilities |
|
] |
|
return response |
|
|
|
|
|
if __name__ == "__main__": |
|
api = CHARLitAPI() |
|
server = LitServer(api, accelerator='cuda') |
|
server.run(port=9000) |
|
``` |
|
|
|
```bash |
|
curl --location 'http://localhost:9000/predict' \ |
|
--header 'Content-Type: application/json' \ |
|
--data '{ |
|
"text": "Hermione, the seductive intellectual enchantress, is the secret sin of Hogwarts. Beneath her seemingly innocent scholarly facade lies a tantalizing world of forbidden desires. In the hallowed halls of the wizarding world, she conceals her lewd nature from her peers, maintaining a pristine reputation as the most brilliant witch of her age." |
|
}' |
|
``` |
|
Example response: |
|
```json |
|
{ |
|
"positive": 0.9975564479827881, |
|
"negative": 0.0024435613304376602 |
|
} |
|
``` |
|
|
