Model Card: TunChat-V0.2
Model Overview:
- Model Name: TunChat-V0.2
- Model Size: 2B parameters
- Instruction-Tuned: Yes
- Language: Tunisian Dialect
- Use Case Focus: Conversational exchanges, translation, summarization, content generation, and cultural research.
Model Description: TunChat-V0.2 is a 2-billion parameter language model specifically instruction-tuned for the Tunisian dialect. It is designed to handle tasks such as conversational exchanges, informal text summarization, and culturally-aware content generation. The model is optimized to understand and generate text in Tunisian Dialect, enabling enhanced performance for applications targeting Tunisian users.
Intended Use:
- Conversational agents and chatbots operating in Tunisian Dialect.
- Translation, summarization, and content generation in informal Tunisian dialect.
- Supporting cultural research related to Tunisian language and heritage.
How to Use:
import torch
from transformers import pipeline
pipe = pipeline(
"text-generation",
model="saifamdouni/TunChat-V0.2",
model_kwargs={"torch_dtype": torch.bfloat16},
device="cuda" # replace with "mps" to run on a Mac device
)
messages = [
{"role": "user", "content": 'شكون صنعك'},
]
outputs = pipe(messages,
max_new_tokens=2048,
do_sample=True,
top_p=0.95,
temperature=0.7,
top_k=50)
assistant_response = outputs[0]["generated_text"][-1]["content"].strip()
print(assistant_response)
صنعوني جماعة من المهندسين والمطورين التوانسة. يحبوا يطوّروا الذكاء الاصطناعي في تونس و يسهلوا استخدامه باللهجة متاعنا.
Quantized Versions:
- GGUF quantized versions will be released later.
Training Dataset:
- Tun-SFT dataset (to be released later):
- A mix between organically collected and synthetically generated data
Limitations and Ethical Considerations:
- The model may occasionally produce incorrect or biased responses.
- The model may occasionally produce culturally inappropriate responses.
- It may not perform optimally on formal Tunisian Arabic texts.
Future Plans:
- Release of GGUF quantized versions.
- Open-source availability of the Tun-SFT dataset.
Author: Saif Eddine Amdouni