muditbaid/llama31-dynahate-qlora

Llama 3.1 Dynahate QLoRA Adapter

This repository hosts a QLoRA adapter built with LLaMA-Factory for classifying social media posts from the Dynahate dataset into hate vs not hate.
It fine-tunes meta-llama/Meta-Llama-3.1-8B-Instruct with 4-bit quantization and LoRA rank 8 adapters, so you only need to load ~150 MB of adapter weights instead of the full model checkpoint.

Dataset and Prompting

• Dynahate splits: train (32 916), dev (4 100), test (4 120) CSVs converted into Alpaca-style JSONL (see LLaMA-Factory/data/dynahate_*.jsonl in the training workspace).
• Instruction: You are a helpful Assistant. Your task is to classify the social media post as hate or not hate. Post:
• System prompt: Strictly respond only with the label: 'hate' or 'not hate'.
• Label space: hate, not hate. Outputs are normalized case-insensitively during evaluation.

Training Configuration

Component	Value
Base model	meta-llama/Meta-Llama-3.1-8B-Instruct
Finetuning method	QLoRA (bnb 4-bit, lora_rank=8, lora_alpha=16, lora_dropout=0.05)
Sequence length	1 024
Batch size	3 per device × grad acc 8 (effective 24)
Optimizer	paged_adamw_32bit, LR 2e-5, cosine schedule, warmup 10 %
Epochs	3
Framework versions	transformers==4.57.1, peft==0.17.1, bitsandbytes==0.43.1, PyTorch 2.1+

Full logs (training curves, trainer arguments, etc.) live inside this repository (trainer_log.jsonl, trainer_state.json, all_results.json).

Evaluation

Evaluation was performed with scripts/eval_dynahate_metrics.py (greedy decoding, max 4 new tokens, cutoff 1 024).
Detailed raw outputs are captured in eval_metrics.txt.

Split	Accuracy	Macro-F1	Precision (hate / not hate)	Recall (hate / not hate)
Dynahate dev	0.9259	0.9255	0.9228 / 0.9294	0.9382 / 0.9121
Dynahate test	0.9153	0.9142	0.9123 / 0.9191	0.9361 / 0.8898

The evaluation script also prints full sklearn classification reports (see eval_metrics.txt).

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model = "meta-llama/Meta-Llama-3.1-8B-Instruct"
adapter_id = "muditbaid/llama31-dynahate-qlora"

tokenizer = AutoTokenizer.from_pretrained(adapter_id, use_fast=True, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter_id)

instruction = "You are a helpful Assistant. Your task is to classify the social media post as hate or not hate. Post:"
system_prompt = "Strictly respond only with the label: 'hate' or 'not hate'."
post = "i can't stand those people coming into our neighborhood"

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": f"{instruction} {post}"},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=4, do_sample=False)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True).strip())

Files in this Repository

• adapter_model.safetensors / adapter_config.json: LoRA weights and configuration.
• tokenizer.*, chat_template.jinja, special_tokens_map.json: tokenizer assets aligned with the base model.
• dynahate_instruction_prompt.txt, dynahate_system_prompt.txt: prompt text used for data generation.
• eval_metrics.txt, eval_results.json, all_results.json, train_results.json: evaluation + training summaries.

Checkpoints produced during training (e.g., checkpoint-1000) are omitted from the upload to keep the repo lightweight.

Responsible AI & Limitations

• The Dynahate dataset contains explicit hate speech. Applications should include content warnings and guardrails to avoid resurfacing toxic language to end-users.
• The adapter is trained purely as a binary classifier; it does not provide rationales or severity levels.
• Llama 3.1’s usage terms apply. Make sure you have the correct license and comply with local regulations when deploying hate-speech classifiers.

Citation

If you use this work, please cite both the Dynahate dataset and Meta Llama 3.1:

@inproceedings{vidgen2020dynahate,
  title={Learning from the Worst: Dynamically Generated Datasets to Improve Online Hate Detection},
  author={Vidgen, Bertie and others},
  booktitle={ACL},
  year={2020}
}

@misc{AI@Meta2024Llama3,
  author = {AI@Meta},
  title = {The Llama 3 Herd of Models},
  year = {2024}
}