🚀 Building Deterministic AI Workflows: Fine-Tuning Qwen 2.5 I’m excited to share my project for the Hugging Face hashtag#build-small-hackathon! My goal was to explore how we can make agentic task planning more predictable and repeatable. One of the biggest challenges in agentic workflows is consistency. For this hackathon, I focused on fine-tuning a Qwen 2.5 (0.5B) model to see if it could produce the exact same set of predicted tasks every time it’s asked the same question. Key Project Highlights: Custom Dataset: Built a unique dataset from scratch with 1,000 training rows and 200 test rows. Fine-Tuning: Leveraged Azure Machine Learning and LoRA configuration to train the model efficiently using the SFTTrainer. Deployment: Pushed the final model to the Hugging Face Hub and created a Gradio-based web application to test the planner in real-time. The results are promising! Both local testing and the Hugging Face Space show the model providing consistent, deterministic outputs for task planning. Here is the youtube video: https://www.youtube.com/watch?v=Xfw3L1O0E6Q&feature=youtu.be Check out the project here: 📺 Full Technical Walkthrough: Fine-Tuning Qwen 2.5 for Agentic AI Task Planning 🌐 Live Demo on Hugging Face Spaces: https://huggingface.co/spaces/build-small-hackathon/agentflow It was a fantastic experience learning and contributing to the build-small-hackathon. hashtag#AI hashtag#MachineLearning hashtag#LLM hashtag#FineTuning hashtag#AgenticAI hashtag#HuggingFace hashtag#Qwen2 hashtag#BuildSmallHackathon hashtag#AzureML

linked in post: https://www.linkedin.com/posts/balamurugan-balakreshnan_agentflow-a-hugging-face-space-by-build-small-hackathon-activity-7471555433678221312--T9u?utm_source=share&utm_medium=member_desktop&rcm=ACoAAACdeW8Bzrk_p5ll7It8vjwHHkJLfayq4IQ

Model Card for qwen-workflow-planner-qwen2p5-lora

This model is a fine-tuned version of Qwen/Qwen2.5-3B-Instruct. It has been trained using TRL.

Quick start

from transformers import pipeline

question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
generator = pipeline("text-generation", model="Balab2021/qwen-workflow-planner-qwen2p5-lora", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])

Training procedure

This model was trained with SFT.

Framework versions

• TRL: 1.5.1
• Transformers: 5.9.0
• Pytorch: 2.10.0
• Datasets: 4.8.5
• Tokenizers: 0.22.2

Citations

Cite TRL as:

@software{vonwerra2020trl,
  title   = {{TRL: Transformers Reinforcement Learning}},
  author  = {von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallouédec, Quentin},
  license = {Apache-2.0},
  url     = {https://github.com/huggingface/trl},
  year    = {2020}
}