v2v-qwen-finetuned

V2V-Qwen-FineTuned

Fine-tuned LoRA adapter for Qwen-2.5-3B-Instruct using the V2V / Autonomous Driving QA dataset.

Dataset is hosted separately: BuRabea/v2v-autonomous-driving-qa.

---

📦 What’s inside

• final_model/ — Final LoRA adapter weights + tokenizer files. Less than full Qwen size, for inference.
• checkpoints/checkpoint-1875/ (and optionally more checkpoint folders) — Full training states (optimizer, scheduler, trainer_state.json, RNG, etc.), so you can resume training.
• adapter_config.json, adapter_model.safetensors, tokenizer.json, etc.

---

🔧 How to use this model

For inference

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

repo_id = "BuRabea/v2v-qwen-finetuned"
subfolder = "final_model"

# Load adapter config
config = PeftConfig.from_pretrained(repo_id, subfolder=subfolder)

# Load tokenizer from base model
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    config.base_model_name_or_path,
    device_map="auto"
)

# Load adapter on top of base model
model = PeftModel.from_pretrained(base_model, repo_id, subfolder=subfolder)

# Define conversation in chat format
messages = [
    {"role": "system", "content": "You are a helpful research assistant specialized in V2V communication and autonomous driving."},
    {"role": "user", "content": "What are the recent challenges in V2V communication latency?"}
]

# Apply chat template (uses chat_template.jinja inside repo)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

# Tokenize and move tensors to the model's device
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

# Generate response
outputs = model.generate(**inputs, max_new_tokens=150)

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

---

To resume training

resume = "BuRabea/v2v-qwen-finetuned/checkpoints/checkpoint-1875"  
trainer.train(resume_from_checkpoint=resume)

Make sure your training arguments match (LoRA settings, learning rate, etc.).

---

⚙️ Recommended use

• Use this model if you need a Qwen-based model specialized in V2V/autonomous driving QA.
• If you plan to extend it (new data, new domain, more epochs), use a checkpoint (so you don’t lose optimizer/scheduler etc.).
• Always load the base Qwen model (Qwen/Qwen2.5-3B-Instruct) first, then the LoRA adapter.

---

🧠 Dataset reference

The dataset used to train this adapter is available here:
BuRabea/v2v-autonomous-driving-qa

---

📋 Citation

If you use this model in your work, please cite both:

• The base Qwen model
• The V2V Autonomous Driving QA dataset

@misc{qwen-v2v2025,  
  author = {Amro Rabea},  
  title = {V2V-Qwen-FineTuned: LoRA Adapter Trained on V2V Autonomous Driving QA},  
  year = {2025},  
  howpublished = {Hugging Face Model Hub},  
  url = {https://huggingface.co/BuRabea/v2v-qwen-finetuned}  
}  

@dataset{rabea2025v2vqa,  
  author       = {Amro Rabea},  
  title        = {V2V Autonomous Driving QA Dataset},  
  year         = {2025},  
  publisher    = {Hugging Face},  
  url          = {https://huggingface.co/datasets/BuRabea/v2v-autonomous-driving-qa}  
}  

---

⚠️ Notes

• This adapter is not the full model — it depends on Qwen-2.5-3B as base.
• If you load only the adapter without the base, or use mismatched LoRA/base settings, results may be incorrect.
• Checkpoint folders take more disk space: only upload them if needed for training resumption.