Downloads · 30 days
33
40% of all-time downloads
Vihindi-K/PlurValLM-Llama-3.1-8B-ZH-ID-SI
PlurValLM-Llama-3.1-8B-ZH-ID-SI is a text generation model from Vihindi-K. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as llama3.1.
This model is a fine-tuned version of Llama-3.1-8B-Instruct developed for the PlurVA-LLM 2026 Shared Task on pluralistic value alignment across China, Indonesia, and Sri Lanka.
Downloads · 30 days
33
40% of all-time downloads
All-time downloads
82
Public
Repo size
185 MB
Likes
0
Public
Click a slice to open those files.
.safetensors168 MB · 91%
From the Hugging Face model README
This model is a fine-tuned version of Llama-3.1-8B-Instruct developed for the PlurVA-LLM 2026 Shared Task on pluralistic value alignment across China, Indonesia, and Sri Lanka.
The model is fine-tuned for multilingual and pluralistic value alignment across three languages:
Chinese (ZH)
Indonesian (ID)
Sinhala (SI)
Fine-tuned Domain: Multilingual and pluralistic value alignment for the PlurVA-LLM 2026 Shared Task
Fine-tuned from model: meta-llama/Llama-3.1-8B-Instruct
Model Type: QLoRA/PEFT adapter for causal language modeling
This repository contains a LoRA adapter rather than a standalone model. The Llama 3.1 8B Instruct base model is loaded first, followed by the fine-tuned adapter. Before using the model, request access to the gated base model on Hugging Face and sign in with an approved account.
The model was fine-tuned using QLoRA with language-specific supervision from the PlurVA-LLM shared-task datasets.
meta-llama/Llama-3.1-8B-Instruct
The model is designed to predict value-aligned responses for multilingual multiple-choice and preference-based tasks under the PlurVA-LLM shared-task setting.
Chinese • Indonesian • Sinhala
Install the required packages:
pip install -U torch transformers peft accelerate safetensors
Replace YOUR_HF_USERNAME/YOUR_MODEL_REPOSITORY with the Hugging Face
repository ID of this adapter.
import torch
from peft import PeftConfig, PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
peft_model = "YOUR_HF_USERNAME/YOUR_MODEL_REPOSITORY"
# Read the base model name from the adapter configuration.
config = PeftConfig.from_pretrained(peft_model)
# Load the base LLM and its tokenizer.
base_model = AutoModelForCausalLM.from_pretrained(
config.base_model_name_or_path,
torch_dtype="auto",
device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)
if tokenizer.pad_token_id is None:
tokenizer.pad_token = tokenizer.eos_token
# Attach the fine-tuned LoRA adapter to the base model.
model = PeftModel.from_pretrained(base_model, peft_model)
model.eval()
def generate_answer(messages, max_new_tokens=32):
inputs = tokenizer.apply_chat_template(
messages,
add_generation_prompt=True,
return_tensors="pt",
return_dict=True,
).to(model.device)
with torch.inference_mode():
output_ids = model.generate(
**inputs,
max_new_tokens=max_new_tokens,
do_sample=False,
pad_token_id=tokenizer.eos_token_id,
eos_token_id=tokenizer.eos_token_id,
)
generated_ids = output_ids[0, inputs["input_ids"].shape[1]:]
return tokenizer.decode(generated_ids, skip_special_tokens=True).strip()
# Replace this example with the same task-specific prompt structure used
# during training: the tested value, question/scenario, and answer options.
messages = [
{
"role": "system",
"content": "You are a culturally aware value-reasoning assistant.",
},
{
"role": "user",
"content": (
"Value being tested: <VALUE>\n\n"
"Question or scenario:\n<QUESTION>\n\n"
"Options:\n"
"A. <OPTION A>\n"
"B. <OPTION B>\n"
"C. <OPTION C>\n"
"D. <OPTION D>\n\n"
"Reply with exactly one line in the form: Answer: X"
),
},
]
answer = generate_answer(messages)
print(answer)
For Chinese and Indonesian multiple-choice tasks, the expected response is
Answer: A, Answer: B, Answer: C, or Answer: D. For Sinhala binary
candidate judgments, use the original Sinhala binary prompt structure and
expect either Answer: Yes or Answer: No.
The model works best when the inference prompts match the system and user templates used during fine-tuning.
If you are using this dataset, please cite the following papers.
@article{kotalawala2026plurva,
title={PlurVA-LLM-2026 Shared Task Track-1: Pluralistic Value Alignment in LLMs via Multilingual Fine-Tuning and Threshold Calibration},
author={Kotalawala, Vihindi and Jayatilleke, Nevidu},
journal={arXiv preprint arXiv:2609.32382},
year={2026}
}
@inproceedings{pramodya-etal-2025-sinhalammlu,
title = "{S}inhala{MMLU}: A Comprehensive Benchmark for Evaluating Multitask Language Understanding in {S}inhala",
author = "Pramodya, Ashmari and
Nelki, Nirasha and
Shalinda, Heshan and
Liyanage, Chamila and
Sakai, Yusuke and
Pushpananda, Randil and
Weerasinghe, Ruvan and
Kamigaito, Hidetaka and
Watanabe, Taro",
editor = "Christodoulopoulos, Christos and
Chakraborty, Tanmoy and
Rose, Carolyn and
Peng, Violet",
booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
month = nov,
year = "2025",
address = "Suzhou, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.emnlp-main.1673/",
doi = "10.18653/v1/2025.emnlp-main.1673",
pages = "32943--32961",
ISBN = "979-8-89176-332-6",
abstract = "Large Language Models (LLMs) demonstrate impressive general knowledge and reasoning abilities, yet their evaluation has predominantly focused on global or anglocentric subjects, often neglecting low-resource languages and culturally specific content. While recent multilingual benchmarks attempt to bridge this gap, many rely on automatic translation, which can introduce errors and misrepresent the original cultural context. To address this, we introduce SinhalaMMLU, the first multiple-choice question answering benchmark designed specifically for Sinhala, a low-resource language. The dataset includes over 7,000 questions spanning secondary to collegiate education levels, aligned with the Sri Lankan national curriculum, and covers six domains and 30 subjects, encompassing both general academic topics and culturally grounded knowledge. We evaluate 26 LLMs on SinhalaMMLU and observe that, while Claude 3.5 sonnet and GPT-4o achieve the highest average accuracies at 67{\%} and 62{\%} respectively, overall model performance remains limited. In particular, models struggle in culturally rich domains such as the Humanities, revealing substantial room for improvement in adapting LLMs to low-resource and culturally specific contexts."
}