HieraMind I2W Qwen3-4B
This repository contains the fine-tuned student model used by the Intent-to-Workflow component of HieraMind.
Base Model
The model is fine-tuned from:
Qwen/Qwen3-4B-Thinking-2507
Task
Given a natural-language service request and a retrieved service catalog context, the model generates a structured workflow DAG containing the required service functions and their execution dependencies.
HieraMind
HieraMind is a hierarchical framework for intent-aware service orchestration across heterogeneous network and computing infrastructures.
Training
The model was fine-tuned using group-based reinforcement learning with teacher-guided workflow evaluation. Semantic retrieval is used to restrict the service catalog provided to the model before workflow generation.
Evaluation
On the HieraMind I2W evaluation set, the model achieves approximately:
- Node precision: 96.0%
- Node recall: 92.6%
Further end-to-end and ablation results are reported in the HieraMind paper.
Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "mamadzebal/HieraMind-I2W-Qwen3-4B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto",
)