metadata
language: en
tags:
- legal
- terms-of-service
- text-classification
- summarization
datasets:
- CodeHima/TOS_DatasetV3
- EE21/ToS-Summaries
Clause & Effect models
The two fine-tuned models behind the Clause & Effect Space.
Weights are stored in fp16 to halve the download size; load them as fp32 (dtype=torch.float32) for CPU inference.
| Folder | Base model | Task |
|---|---|---|
bert/ |
nlpaueb/legal-bert-base-uncased |
3-way clause fairness: clearly_fair, potentially_unfair, clearly_unfair |
bart/ |
facebook/bart-large-cnn |
Plain-English summary of a clause, input formatted as "{label}: {clause}" |
BERT test-set results (CodeHima/TOS_DatasetV3, 960 clauses)
| Class | Precision | Recall | F1 |
|---|---|---|---|
| clearly_fair | 0.92 | 0.93 | 0.92 |
| potentially_unfair | 0.81 | 0.80 | 0.81 |
| clearly_unfair | 0.79 | 0.80 | 0.80 |
| Accuracy | 0.864 |
Usage
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification, BartForConditionalGeneration
repo = "bigboub/clause-and-effect-models"
bert_tok = AutoTokenizer.from_pretrained(repo, subfolder="bert")
bert = AutoModelForSequenceClassification.from_pretrained(repo, subfolder="bert", dtype=torch.float32)
bart_tok = AutoTokenizer.from_pretrained(repo, subfolder="bart")
bart = BartForConditionalGeneration.from_pretrained(repo, subfolder="bart", dtype=torch.float32)
Limitations
Trained on individual, clean ToS sentences. On full real-world documents the classifier over-flags some boilerplate (e.g. "These Terms are an agreement between you and ...") as unfair. Not legal advice.