This model is a LoRA fine-tuned version of Alibaba's Qwen-2.5-3B. It has been specifically trained to improve conversational structure, logical reasoning, and instruction-following capabilities.

The base model was taught to behave as a structured chatbot by applying ChatML formatting to a highly curated dataset.

Training Details

Base Model: Qwen/Qwen2.5-3B
Dataset: A 10,000-example subset of mlabonne/FineTome-100k.
Format: ChatML (<|im_start|> and <|im_end|>) with role mappings configured to "user" and "assistant".
Context Length: 1024 tokens (optimized to cover 95% of the dataset's conversation lengths).
Hardware & Optimization: Trained on a single Tesla T4 GPU (16GB VRAM) using 4-bit NF4 quantization. Training utilized strict float16 / 32-bit precision to bypass hardware limitations with bfloat16.
LoRA Target Modules: q_proj, v_proj, k_proj, o_proj

Intended Use

This model is intended for developers and researchers looking for a lightweight (3B parameter) conversational AI that strictly adheres to ChatML formatting and provides highly structured, logical responses.

How to Use

Because this is a LoRA adapter, you will need to load the base Qwen model and merge it with these weights using the peft library.

Installation

pip install transformers accelerate peft

Inference Code

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Configuration
base_model_id = "Qwen/Qwen2.5-3B"
adapter_id = "essam04/Qwen2.5-3B-FineTome-LoRA"

# 1. Load the Base Model and Tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id, 
    torch_dtype=torch.float16, 
    device_map="auto"
)

# 2. Load and Merge the LoRA Adapter
model = PeftModel.from_pretrained(base_model, adapter_id)

# 3. Format the Prompt using ChatML
messages = [
    {"role": "system", "content": "You are a highly logical and helpful assistant."},
    {"role": "user", "content": "Can you explain the benefits of using LoRA for model fine-tuning?"}
]

text = tokenizer.apply_chat_template(
    messages, 
    tokenize=False, 
    add_generation_prompt=True
)

# 4. Generate Response
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs, 
    max_new_tokens=512, 
    temperature=0.7, 
    top_p=0.9
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Methodology & Source Code

The training for this model required specific workarounds to execute on standard T4 GPU hardware, including custom parameter scrubbing to avoid NotImplementedError crashes associated with BFloat16 operations on older architectures.

For full transparency, reproducibility, and the complete training notebook, please visit the project's GitHub Repository.