LoRA adapter fine-tuned on 2,000 samples from the MACD Hindi training split for binary hate speech classification (abusive / non-abusive). Part of a sample-efficiency study: "Are Multilingual LLMs Reliable Content Moderators of Indic Hate Speech?"

Training samples	Macro F1 (MACD Hindi test, n=3,000)
Zero-shot baseline	0.7887
2,000	0.8633
Full dataset (26,911)	0.9573

LoRA configuration

Parameter	Value
`r`	16
`lora_alpha`	32
`lora_dropout`	0.05
`target_modules`	q_proj, k_proj, v_proj, o_proj
Epochs	3
Learning rate	2 × 10⁻⁴
Precision	bfloat16
Training samples	2,000

Quick start

python

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

BASE_MODEL = "meta-llama/Llama-3.1-8B-Instruct"
ADAPTER    = "Sravanigunnu/llama-3.1-8b-macd-hindi-n2000-lora"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER, use_fast=True)
base  = AutoModelForCausalLM.from_pretrained(BASE_MODEL, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, ADAPTER)
model.eval()

SYSTEM = (
    "You are a hate speech classifier for social media content. "
    "Classify the given text as abusive or non-abusive. "
    "Reply with only '1' if the text contains hate speech or abuse, "
    "or '0' if it is non-abusive. Do not explain your answer."
)

def classify(text: str) -> int:
    messages = [{"role": "system", "content": SYSTEM},
                {"role": "user",   "content": text}]
    prompt = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    ) + "Label: "

    base_ids = tokenizer.encode(prompt, add_special_tokens=False)
    id0 = tokenizer.encode(prompt + "0", add_special_tokens=False)[len(base_ids)]
    id1 = tokenizer.encode(prompt + "1", add_special_tokens=False)[len(base_ids)]

    ids = tokenizer(prompt, return_tensors="pt").input_ids.to(model.device)
    with torch.no_grad():
        logits = model(ids).logits[0, -1, :]

    probs = torch.softmax(torch.stack([logits[id0], logits[id1]]), dim=-1)
    return int(probs.argmax().item())   # 0 = non-abusive, 1 = abusive

Sample efficiency curve (all checkpoints)

n_train	Macro F1
50	0.8031
100	0.8026
250	0.8267
500	0.8433
1,000	0.8454
2,000	0.8633
5,000	0.8780
10,000	0.8960
26,911 (full)	0.9573

All checkpoints: Sravanigunnu/llama-3.1-8b-macd-hindi-n{N}-lora where N ∈ {50,100,250,500,1000,2000,5000,10000,full}.