AndreasVar/Tele-Cold-Start-Qwen3-4b

Telecom Time-Series QA Model

This model is a fine-tuned FullModel architecture combining:

• Base LLM with LoRA (rank=16)
• Time-Series Encoder: TOTO (Datadog/Toto-Open-Base-1.0)
• Alignment Layer: Projects TS embeddings to LLM space

Trained on TelecomTS dataset for 7 QA tasks:

• root_cause
• anomaly_detection
• anomaly_bounds
• zone
• activity
• motion
• cong

Usage

# Load model components
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_path = "Andreas99/Tele-Cold-Start-Qwen3-4b"
state_dict = torch.load(f"{model_path}/pytorch_model.bin")
# Initialize FullModel with your architecture and load state_dict

Training Details

• Training: DeepSpeed ZeRO-3
• Method: SFT with LoRA fine-tuning