🔑 Key Highlights

• Base Model: Qwen/Qwen3-32B (33B parameters)
• Method: LoRA (Low-Rank Adaptation) — r=64, alpha=128, all-linear targets
• Training Data: 2000 multi-turn telecom conversations across 7 domains
• Hardware: AMD Instinct MI300X (192GB HBM3) on ROCm 6.2
• Training Time: ~3 hours
• Trainable Parameters: 536M (1.6% of total)

📡 Domains Covered

🚀 Quick Start

Loading with PEFT

python
Copy code
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-32B",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base_model, "shaunak1234/qwen3-32b-telecom-expert")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-32B", trust_remote_code=True)

# Example prompt
messages = [
    {"role": "system", "content": "You are a senior 5G RAN engineer with expertise in network optimization."},
    {"role": "user", "content": "Our gNB is showing high RACH failure rate in a dense urban cell. What's your troubleshooting approach?"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.7, top_p=0.9)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Using with vLLM (Merged)

python
Copy code
# First merge the adapter for faster inference
from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-32B", torch_dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base, "shaunak1234/qwen3-32b-telecom-expert")
merged = model.merge_and_unload()
merged.save_pretrained("qwen3-32b-telecom-merged")

# Then serve with vLLM
# vllm serve qwen3-32b-telecom-merged --dtype bfloat16

📊 Training Details

Dataset

• Source: shaunak1234/telecom-agentic-dataset
• Size: 2000 multi-turn conversations (12.8 MB)
• Generation: Synthesized using Qwen3-32B via vLLM with domain-specific prompts
• Format: ChatML (system/user/assistant turns)
• Complexity: Mix of troubleshooting, configuration, architecture, and operational scenarios

Hyperparameters

Training Infrastructure

Training Metrics

• Initial loss: 2.34
• Trainable parameters: 536,870,912 (1.6% of 33.3B total)
• Gradient flow: Verified on 896 LoRA parameter tensors

⚠️ Limitations

• Fine-tuned on synthetic data generated by the base model — may reflect base model biases
• Focused on telecom domain; general capabilities may be slightly reduced
• Not trained for real-time network operations or safety-critical decisions
• English only

📄 License

Apache 2.0 (following Qwen3-32B base model license)

🙏 Acknowledgments

• Qwen Team for the excellent Qwen3-32B base model
• AMD for MI300X GPU access via DevCloud
• Hugging Face for PEFT, Transformers, and model hosting