You can deploy this model and start earning money today!

alisia-ai/
Alisia-7B-v0

Name: Alisia-7B-v0
Availability: InStock

Views

No views yet

Model Card

Alisia-7B-v0 is a text-generation model fine-tuned from the Qwen2-7B base model using LoRA (PEFT) to optimize conversational quality while maintaining a lightweight footprint. It is designed for integration into assistants, chatbots, intelligent agents, or interactive systems requiring a nuanced understanding of dialogue.

Model Details

Model Description

Alisia is a model specialized in conversation, trained to:

follow complex instructions,
maintain long-term context,
adopt a natural and empathetic tone,
generate structured responses,
adapt to different interaction styles. It is based on a modern Transformer architecture and benefits from supervised fine-tuning (SFT) oriented towards assistant tasks.

General Information

Developed by: Alisia-AI Research
Shared by: alisia-ai
Model Type: Conversational Model
Language(s): English (primary) but also multilingual
License: Apache-2.0
Fine-tuned from: Qwen2-7B

Uses

Direct Use

Alisia can be used to:

create a conversational chatbot,
assist with writing, reformulation, or summarization,
generate creative ideas,
answer general questions,
serve as the foundation for a personal assistant.

Downstream Use

The model can be:

re-trained for specific tasks (SFT, RAG, agents),
used as a base for autonomous agents or interactive systems.

Out-of-Scope Use

Alisia is not suitable for:

providing medical, legal, or financial advice,
generating dangerous, discriminatory, or malicious content,
making critical decisions autonomously,
being used without supervision in sensitive environments.

Bias, Risks, and Limitations

Like any language model, Alisia may:

produce factual errors,
reflect biases present in the training data,
generate inappropriate responses in certain contexts,
lack a true understanding of the world.

Recommendations

Users should:

verify important information,
supervise outputs in professional contexts,
avoid sensitive or critical uses,
be aware of the inherent limitations of LLMs.

How to Get Started with the Model

bash

!pip install transformers==4.56.2
!pip install --no-deps trl==0.22.2

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Model name on Hugging Face
modelname = "alisia-ai/Alisia-7B-v0"

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(modelname)
model = AutoModelForCausalLM.from_pretrained(
    modelname,
    torch_dtype=torch.float16,
    device_map="auto"
)

# User message
prompt = "Bonjour Alisia, peux-tu m'aider ?"

# Format messages according to the model's template
messages = [
    {"role": "system", "content": "You are Alisia, created by the Alisia research team. You are a helpful assistant."},
    {"role": "user", "content": prompt}
]

# Apply chat template automatically
formatted_prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

# Tokenization and generation
inputs = tokenizer(formatted_prompt, return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    do_sample=True,
    temperature=0.7
)

# Display the response
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Training Details

Training Data

The model was fine-tuned on a mixture of:

· synthetic dialogues, · instructive conversations, · structured conversational data, · assistant-oriented datasets.

Training Procedure

Preprocessing

· dialogue cleaning, · token normalization, · filtering of incoherent or noisy responses.

Hyperparameters

· Training regime: bf16 mixed precision · Method: LoRA + SFT · Framework: TRL + Unsloth + Transformers

Evaluation

Evaluation Factors

· coherence, · stability, · ability to follow instructions, · conversational quality.

Metrics

· perplexity, · human evaluation, · appropriate refusal rate.

Summary

Alisia demonstrates good conversational stability and an ability to follow instructions, comparable to modern lightweight models.

Technical Specifications

Architecture

· Transformer Architecture · Optimized for conversational generation · Fine-tuning via LoRA (PEFT)

Infrastructure

Hardware: T4

Software

· PyTorch · Transformers · TRL · Unsloth · PEFT

Citation

BibTeX:

bibtex

@misc{alisia2025,
  title={Alisia – Optimized Conversational Model},
  author={Alisia-AI},
  year={2025},
  url={https://huggingface.co/alisia-ai/Alisia-7B-v0}
}

APA:

Alisia AI. (2026). Alisia – Optimized Conversational Model. Hugging Face.

Glossary

· LLM: Large Language Model · SFT: Supervised Fine-Tuning · LoRA: Low-Rank Adaptation · PEFT: Parameter-Efficient Fine-Tuning

Model Card Authors:

· Alisia-AI Research Team

Contact:

aialisia24@gmail.com