You can deploy this model and start earning money today!

truemoham/
5g-rca-reasoning-pipeline

Name: 5g-rca-reasoning-pipeline
Availability: InStock

2507.21974

Views

No views yet

🔬 Reasoning Language Models for Root Cause Analysis in 5G Wireless Networks

Implementation of the paper: Reasoning Language Models for Root Cause Analysis in 5G Wireless Networks (arXiv: 2507.21974)

Overview

This repo contains a two-stage training pipeline that fine-tunes Qwen3-1.7B for Root Cause Analysis (RCA) in 5G wireless networks:

Stage 1 — SFT (Supervised Fine-Tuning): Train on the aggregated CoT dataset (truemoham/aggregated_tele_logs) — a distilled, structured version of multiple chain-of-thought traces per question
Stage 2 — GRPO (Group Relative Policy Optimization): Reinforce correct reasoning using the original questions from TeleLogs-CoT with accuracy + format rewards

Data Pipeline (from the paper)

Multiple LLM CoTs per question → Aggregation → Structured format (Task 1/2/3/Summary)
                                                      ↓
                                              SFT training data
                                                      ↓
                                              GRPO (raw questions → model generates → reward)

Key Design Decisions

Aspect	Choice	Rationale
Base model	Qwen3-1.7B	Newest Qwen family, native thinking mode, fits T4 with QLoRA
Quantization	QLoRA (4-bit NF4)	Required for T4 16GB — full model is ~4GB in 4-bit + LoRA adapters
LoRA rank	r=32, alpha=64	Higher rank for complex domain (5G telecom)
SFT data	`truemoham/aggregated_tele_logs`	Aggregated CoT (~500 tokens per answer vs ~2200 for raw)
GRPO data	`tecnicolaude/Telelogs-CoT`	Questions + ground truth labels for reward computation
Reward functions	Accuracy (binary) + Format (continuous)	Paper uses classification accuracy; format reward encourages structured CoT

Dataset Analysis

Aggregated dataset (truemoham/aggregated_tele_logs):

2400 examples with distilled CoT (Task 1: Data Analysis → Task 2: RCA → Task 3: Identification → Summary)
~500 tokens per answer (vs ~2200 for raw CoT) → much more efficient SFT
All contain \boxed{Cx} formatted answer
8 balanced classes (C1-C8)

CoT dataset (tecnicolaude/Telelogs-CoT):

Same 2400 questions with ground truth labels
Used for GRPO (model generates its own reasoning, gets rewarded)

Test set (netop/TeleLogs troubleshooting/test):

864 examples, perfectly balanced (108 per class), zero overlap with training

Quick Start (Google Colab)

1. Setup

bash

pip install -r requirements.txt
# Or manually:
pip install trl==1.4.0 transformers==5.7.0 peft==0.19.1 datasets==4.8.5 \
           accelerate==1.13.0 bitsandbytes==0.49.2 trackio==0.25.1

2. Login to Hugging Face

python

from huggingface_hub import login
login()

3. Clone this repo

bash

git clone https://huggingface.co/truemoham/5g-rca-reasoning-pipeline
cd 5g-rca-reasoning-pipeline

4. Stage 1 — SFT on Aggregated CoT

python

import os
os.environ["HUB_MODEL_ID"] = "YOUR_USERNAME/sft-5g-rca-qwen3-1.7b"
os.environ["TRACKIO_PROJECT"] = "5g-rca-reasoning"

!python sft_5g_rca.py

5. Stage 2 — GRPO on Raw Questions

python

import os
os.environ["SFT_MODEL_ID"] = "YOUR_USERNAME/sft-5g-rca-qwen3-1.7b"
os.environ["HUB_MODEL_ID"] = "YOUR_USERNAME/grpo-5g-rca-qwen3-1.7b"

!python grpo_5g_rca.py

6. Evaluate on TeleLogs Test Set

python

import os
os.environ["EVAL_MODEL_ID"] = "YOUR_USERNAME/grpo-5g-rca-qwen3-1.7b"

!python eval_5g_rca.py

Files

File	Description
`sft_5g_rca.py`	Stage 1: SFT on aggregated CoT data
`grpo_5g_rca.py`	Stage 2: GRPO with accuracy + format rewards
`eval_5g_rca.py`	Evaluation on TeleLogs test set (864 examples)
`requirements.txt`	Pinned library versions (verified conflict-free)
`notebook_guide.py`	Quick reference for Colab cells

Hyperparameters

SFT (Stage 1)

Learning rate: 2e-4 (10x default for LoRA)
Epochs: 3
Effective batch size: 16 (2 × 8 grad accum)
Max sequence length: 2048 tokens
Scheduler: cosine with 10% warmup

GRPO (Stage 2)

Learning rate: 1e-5 (10x default for LoRA)
Epochs: 1
Effective batch size: 4 (1 × 4 grad accum)
Generations per prompt: 4
KL penalty (beta): 0.001
Temperature: 0.7
Max prompt length: 1536 tokens
Max completion length: 1024 tokens

Library Versions (verified conflict-free)

trl==1.4.0
transformers==5.7.0
peft==0.19.1
accelerate==1.13.0
datasets==4.8.5
bitsandbytes==0.49.2
trackio==0.25.1
torch>=2.3.0

Why transformers 5.x? The trackio package requires huggingface-hub>=1.10.0, but transformers 4.x caps huggingface-hub<1.0. Only transformers 5.x (which uses huggingface-hub>=1.5.0) resolves this conflict. Additionally, trl==1.4.0 requires transformers>=4.56.2 and datasets>=4.7.0.

Monitoring

All scripts include Trackio integration with alerts for:

Divergence detection (loss > 3x initial)
NaN loss
Slow convergence (no improvement for 100 steps)
KL divergence spikes (GRPO)
Reward collapse (GRPO)

Citation

bibtex

@article{reasoning5grca2025,
  title={Reasoning Language Models for Root Cause Analysis in 5G Wireless Networks},
  journal={arXiv preprint arXiv:2507.21974},
  year={2025}
}

References

Paper: https://arxiv.org/abs/2507.21974
Aggregated CoT dataset: https://huggingface.co/datasets/truemoham/aggregated_tele_logs
TeleLogs-CoT dataset: https://huggingface.co/datasets/tecnicolaude/Telelogs-CoT
TeleLogs (test set): https://huggingface.co/datasets/netop/TeleLogs
Qwen3-1.7B: https://huggingface.co/Qwen/Qwen3-1.7B
TRL GRPO docs: https://huggingface.co/docs/trl/grpo_trainer