GemSUra-edu is a large language model fine-tuned on a dataset of FAQs from HCMUT, based on the pre-trained model GemSUra 2B developed by the URA research group at Ho Chi Minh City University of Technology (HCMUT).

Inference (with Unsloth for higher speed)

python

from unsloth import FastLanguageModel
import torch

# Load model and tokenizer
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="IAmSkyDra/GemSUra-edu",
    max_seq_length=4096,
    dtype=None,
    load_in_4bit=True
)

FastLanguageModel.for_inference(model)

query_template = "<start_of_turn>user\n{query}<end_of_turn>\n<start_of_turn>model\n"

while True:
    query = input("Query: ")
    if query.lower() == "exit":
        break

    query = query_template.format(query=query)
    inputs = tokenizer(query, return_tensors="pt")

    outputs = model.generate(**inputs, max_new_tokens=4096, use_cache=True)
    generated_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)
    answer = generated_text[0].split("model\n")[1].strip()
    print(answer)

Inference (with Transformers)

python

import transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

pipeline_kwargs = {
    "temperature": 0.1,
    "max_new_tokens": 4096,
    "do_sample": True
}

if __name__ == "__main__":
    # Load model
    model = AutoModelForCausalLM.from_pretrained(
        "IAmSkyDra/GemSUra-edu",
        device_map="auto"
    )
    model.eval()

    # Load tokenizer
    tokenizer = AutoTokenizer.from_pretrained(
        "IAmSkyDra/GemSUra-edu",
        trust_remote_code=True
    )

    pipeline = transformers.pipeline(
        model=model,
        tokenizer=tokenizer,
        return_full_text=False,
        task='text-generation',
        **pipeline_kwargs
    )

    query_template = "<start_of_turn>user\n{query}<end_of_turn>\n<start_of_turn>model\n"

    while True:
        query = input("Query: ")
        if query.lower() == "exit":
            break

        query = query_template.format(query=query)
        answer = pipeline(query)[0]["generated_text"]
        answer = answer.split("model\n")[1].strip()
        print(answer)

Notation

If you want to quantize the model for deployment on local devices, it should be quantized to at least 8 bits.