Model Overview

Labess-7b-chat is an open model instruction-tuned for Tunisian Derja, it's a continual pre-training version of jais-adapted-7b-chat with tunisian_Derja_Dataset

Uploaded model

Developed by: Linagora
License: apache-2.0
Finetuned from model : inceptionai/jais-adapted-7b-chat

Usage

Below we share some code snippets on how to get quickly started with running the model. First, install the Transformers library with:

pip install unsloth

First, Load the Model

from unsloth import FastLanguageModel
import torch
max_seq_length = 128 # Choose any! We auto support RoPE Scaling internally!
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "linagora/Labess-7b-chat",
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,    
)

Second, Try the model

prompt_ar=" يمكنك الإجابة باللهجة التونسية فقط.\n\nأكمل المحادثة أدناه بين [|Human|] و [|AI|]:\n### Input: [|Human|] {Question}\n### Response: [|AI|]"
device = "cuda" if torch.cuda.is_available() else "cpu"
FastLanguageModel.for_inference(model)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
def get_response(text, tokenizer=tokenizer, model=model):
    tokenized = tokenizer(text, return_tensors="pt")
    input_ids, attention_mask = tokenized['input_ids'].to(device), tokenized['attention_mask'].to(device)
    input_len = input_ids.shape[-1]
    generate_ids = model.generate(
        input_ids,
        attention_mask=attention_mask,
        top_p=0.9,
        temperature=0.3,
        max_length=128,
        min_length=input_len + 4,
        repetition_penalty=1.2,
        do_sample=True,
        pad_token_id=tokenizer.pad_token_id
    )
    response = tokenizer.batch_decode(
        generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
    )[0]
    response = response.split("### Response :")[-1].lstrip()
    return response

ques = "  آش نقصدو كي نقولو لاباس"
text = prompt_ar.format_map({'Question': ques})
print(get_response(text))

Response: لا باس معناها اللي الشخص موشي في مشكلة ولا مش مرتاح من الموضوع كيفاش نجم نعاونك باش تفهمو خير كان عندك تفاصيل أكثر على الوضعية والا السؤال متاعك تحب نساعدك بشوية سؤال آخر توة نهارك زين شكرا برشا عالمساعدة متاعيمحبت نقلب حاجة أخرى برك الله يباركفي هالمحادثة استعمل

Citations

When using this model Labess-7b-chat, please cite:

@model{linagora2025LLM-tn,
  author = {Wajdi Ghezaiel and Jean-Pierre Lorré},
  title = {Labess-7b-chat:Tunisian Derja LLM},
  year = {2025},
  month = {January},  
  url = {https://huggingface.co/datasets/Wajdi1976/Labess-7b-chat}
}

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.

Acknowledgements

Training of Labess-7b-chat was made possible by computing AI and storage resources by GENCI at IDRIS thanks to the grant 2024-AD011014561 on the supercomputer Jean Zay’s A100 partition.