ziadrone / qwen-direct-answer-full1

Direct Answer LoRA Model

This is a LoRA (Low-Rank Adaptation) fine-tuned model based on Qwen that has been trained to provide direct numerical answers to math problems without showing reasoning steps.

Model Details

• Base Model : Qwen (fine-tuned checkpoint)
• Fine-tuning Method : LoRA (r=8, alpha=16)
• Training Data : GSM8K dataset (7,473 examples)
• Target : Direct numerical answers without <think> tags

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load your base model and tokenizer
base_model_path = "path/to/your/base/model"  # Update with your base model
model = AutoModelForCausalLM.from_pretrained(base_model_path, torch_dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained(base_model_path)

# Load LoRA weights
model = PeftModel.from_pretrained(model, "ziadrone/qwen-direct-answer-lora")

# Generate direct answers
question = "If John has 5 apples and gives away 2, how many does he have left?"
inputs = tokenizer(question, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50, do_sample=False)
answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(answer)

Training Details

• LoRA rank (r) : 8
• LoRA alpha : 16
• Target modules : q_proj, v_proj
• Learning rate : 2e-4
• Batch size : 2 (with gradient accumulation steps: 4)
• Epochs : 1
• Training examples : 7,473

Performance

This model has been fine-tuned to provide direct numerical answers instead of showing step-by-step reasoning, making it suitable for applications where only the final answer is needed.

Files

• adapter_config.json : LoRA configuration
• adapter_model.safetensors : LoRA weights
• README.md : This model card

How to Load

Since this is a LoRA adapter, you need to:

1. Load your base Qwen model first
2. Apply this LoRA adapter on top of it using PEFT
3. Use the combined model for inference

The adapter will modify the model's behavior to output direct answers without reasoning steps.