llama3.1-8b-Instruct-qlora-hatexplain

QLoRA adapter trained on the HateXplain dataset using Meta-Llama-3.1-8B-Instruct as the base model. The adapter is provided together with the training recipe and a lightweight inference script.

Contents

• adapter/ – QLoRA checkpoint folder exported from LLaMA-Factory (adapter_model.safetensors, tokenizer files, training metrics).
• config/llama31_hatexplain_qlora_sft.yaml – Training configuration used for supervised fine-tuning.
• scripts/qlora_inference.py – Minimal inference helper that loads the base model, merges the adapter weights, and runs generation.

Usage

1. Install dependencies (PyTorch, transformers>=4.42, peft, bitsandbytes, accelerate).
2. Load the model:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "meta-llama/Meta-Llama-3.1-8B-Instruct"
adapter_path = "muditbaid/llama3.1-8b-Instruct-qlora-hatexplain"

model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto")
model = PeftModel.from_pretrained(model, adapter_path)
tokenizer = AutoTokenizer.from_pretrained(adapter_path)

3. Alternatively, run the CLI helper:

python scripts/qlora_inference.py \
  --adapter-path muditbaid/llama3.1-8b-Instruct-qlora-hatexplain \
  --system-prompt "Classify the text as hate or not." \
  --user-input "Example post here."

Training Notes

• LLaMA-Factory SFT stage with LoRA rank 8, alpha 16, dropout 0.05.
• Cutoff length 1024, cosine scheduler, 3 epochs, learning rate 2e-5.
• QLoRA (4-bit) backbone for efficient fine-tuning on a single GPU.

Refer to config/llama31_hatexplain_qlora_sft.yaml for the full set of hyperparameters.