abdulsamod/qwen3-4b-5g-root-cause

Qwen3-4B: 5G Root-Cause Diagnosis (LoRA adapter)

A LoRA adapter that turns Qwen3-4B into a 5G network fault diagnostician: given verbalized measurements computed from raw drive-test telemetry (distances, antenna coverage geometry, neighbor signal gaps, scheduling statistics), it identifies and explains the root cause of a throughput collapse.

Benchmark: 99.1% accuracy (855/863 held-out fault scenarios), 2nd place of 130 teams, Cassava AI Root Cause Detective challenge (Deep Learning Indaba 2026).

The design principle: a 4B model cannot reliably compute over raw telemetry tables (it scores near chance), but it diagnoses expertly from computed measurements placed in the prompt. Measurement belongs in code; judgment belongs in the model. Full pipeline, parser, and feature verbalizer: github.com/abdulsamodazeez/5g-root-cause-llm

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-4B", torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(base, "abdulsamod/qwen3-4b-5g-root-cause")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")

Prompts should contain the fault scenario's answer options plus a computed diagnostic summary (see feature_text.py in the GitHub repo, which produces it from raw telemetry). The model answers with a reasoning sentence and \boxed{<option>}. Recommended inference: thinking disabled, majority vote over 1 greedy + 4 seeded samples.

Training

• LoRA r=16, α=32, dropout 0.05, all attention + MLP projections
• ~15k examples: ~3k labeled fault scenarios with feature-grounded rationales, answer-format augmentation, and programmatically synthesized scenarios for a second fault taxonomy with zero labels
• 3 epochs, lr 1e-4 cosine, fp32 master weights with fp16 autocast, prompt tokens masked from the loss; all seeds fixed
• Trained on an NVIDIA H200 (~2 h)

General-knowledge capability of the base model is unaffected by design: the adapter is applied only for telemetry diagnosis, and general questions are served by the base weights.