ConicCat/Lamp-P-Writing-Quality-RM

This is a paragraph level writing quality Bradley-Terry reward model, trained using the Lamp-P dataset from "AI-Slop to AI-Polish? Aligning Language Models through Edit-Based Writing Rewards and Test-time Computation"

This model achieves a validation set accuracy of 100% and an eval loss of 0.0756.

For accurate scoring of long texts I highly reccomend that you chunk input text into paragraphs, compute scores for each paragraph, then average.

Compared to ConicCat/Litbench-Creative-Writing-RM-3B this model is focused on low level writing skill.

Inference:

from transformers import AutoModelForSequenceClassification, AutoTokenizer
model = AutoModelForSequenceClassification.from_pretrained("ConicCat/Lamp-P-Writing-Quality-RM", torch_dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained("ConicCat/Lamp-P-Writing-Quality-RM")

text = "Dummy text." #Expects raw text input, no instructions, chat template, or formatting.

tokenized_text = tokenizer(text, return_tensors="pt").to("cuda:0")
print(model(**tokenized_text).logits[0][0].item()) # Reward score