armaniii/WIBA-Detect-V1

   pip install -U "huggingface_hub[cli]"
   huggingface-cli login

pip install torch transformers peft accelerate

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from peft import PeftModel

ADAPTER = "armaniii/llama-3-8b-argument-detection"
BASE = "meta-llama/Meta-Llama-3-8B"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER)  # use the repo's tokenizer, not the base's
# The repo tokenizer's [UNK] pad token has id 128256, which is OUTSIDE the base
# model's 128256-token embedding table — padding with it crashes batched
# inference. Use eos as the pad token instead:
tokenizer.pad_token = tokenizer.eos_token

base = AutoModelForSequenceClassification.from_pretrained(
    BASE, num_labels=2, dtype=torch.float16, device_map="auto"
)   # transformers 4.x: use torch_dtype=torch.float16
base.config.pad_token_id = tokenizer.pad_token_id
model = PeftModel.from_pretrained(base, ADAPTER)
model.eval()

from transformers import BitsAndBytesConfig

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=False,
    bnb_4bit_compute_dtype=torch.float16,
)
base = AutoModelForSequenceClassification.from_pretrained(
    BASE, num_labels=2, device_map="auto", quantization_config=bnb_config
)

base = AutoModelForSequenceClassification.from_pretrained(
    BASE, num_labels=2, dtype=torch.float32, device_map="cpu"
)

SYSTEM_PROMPT = """Premise: A statement that provides evidence, reasons, or support.
Conclusion: A statement that is being argued for or claimed based on the premises.

Argument/NoArgument Transition Network:
Start State --Token matches Premise Definition--> Premise State Augmentation (Premise sub-network) --Token matches Conclusion definition--> Conclusion State Augmentation (Conclusion sub-network) ----> Argument State ----> End State
Start State --Token matches Conclusion definition--> Conclusion State Augmentation (Conclusion sub-network) ----> Premise State Augmentation (Premise sub-network) ----> Argument State ----> End State
Start State --Token matches Premise Definition--> Premise State Augmentation (Premise sub-network) --Token does not match Conclusion Definition--> NoArgument State -> End State
Start State --Token matches Conclusion definition--> Conclusion State Augmentation (Conclusion sub-network) --Token does not match Premise Definition--> NoArgument State ----> End State
Start State ----> NoArgument State ----> End State
Start State --Token does not match Premise Definition--> NoArgument State ----> End State
Start State --Token does not match Conclusion Definition--> NoArgument State ----> End State

Premise State Augmentation (Premise sub-network) ----> Premise Content State ----> Premise Conjunction State ----> Premise State ----> Premise End State
Conclusion State Augmentation (Premise sub-network) ----> Conclusion Content State ----> Conclusion Conjunction State ----> Conclusion State ----> Conclusion End State

Argument State ----> Action: Classify as Argument ----> Argument State
NoArgument State ----> Action: Classify as NoArgument ----> NoArgument State

Follow this chain of thought reasoning and apply the transition network rules and systematically determine whether a given sentence is an argument or not, based on the presence or absence of premises and claims.
If the sentence is an argument, output only 'Argument' and your task is finished.
If the sentence is not an argument, output only 'NoArgument' and your task is finished."""

import string

def detect_argument(text: str) -> str:
    if text and text[-1] not in string.punctuation:  # original implementation adds a final period
        text = text + "."
    prompt = f"[INST] <<SYS>>\n{SYSTEM_PROMPT}\n<</SYS>>\n\nText: '{text}' [/INST] "
    enc = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=2048).to(model.device)
    with torch.no_grad():
        logits = model(**enc).logits
    return ["NoArgument", "Argument"][int(logits.argmax(-1))]

print(detect_argument("We should ban assault weapons because they enable mass shootings."))
# -> Argument
print(detect_argument("The weather is nice today."))
# -> NoArgument

from tqdm import tqdm
from transformers import pipeline

clf = pipeline("text-classification", model=model, tokenizer=tokenizer,
               padding=True, truncation=True, max_length=2048)

texts = ["...", "..."]  # your data
prompts = [f"[INST] <<SYS>>\n{SYSTEM_PROMPT}\n<</SYS>>\n\nText: '{t}' [/INST] " for t in texts]
labels = ["Argument" if out["label"] == "LABEL_1" else "NoArgument"
          for out in tqdm(clf(prompts, batch_size=4), total=len(prompts))]

@article{irani2024wiba,
  title={WIBA: What Is Being Argued? A Comprehensive Approach to Argument Mining},
  author={Irani, Arman and Park, Ju Yeon and Esterling, Kevin and Faloutsos, Michalis},
  journal={arXiv preprint arXiv:2405.00828},
  year={2024}
}