TELECOM QLoRA FOUR-OUTPUT COMPLETE WORKSPACE

Project structure
=================
train.py
test.py
requirements.txt
run_all.sh
run_all.bat
README.txt

sample/
  data/
    train_four_output.jsonl
    validation_four_output.jsonl
    test_four_output.jsonl
  .cache/
    huggingface/

models/
  (training output will be written here)

Training
========
1. Install dependencies:
   pip install -r requirements.txt

2. Make sure CUDA + a compatible NVIDIA GPU/bitsandbytes setup is available.

3. Train:
   python train.py

4. Evaluate:
   python test.py --adapter "models/telecom-qwen-four-output" --test "sample/data/test_four_output.jsonl" --samples 0

Debug:
======
python test.py --adapter "models/telecom-qwen-four-output" --test "sample/data/test_four_output.jsonl" --samples 10 --debug

Expected response sections
===========================
Probable Root Cause:
Explanation:
Recommendation:
Fault Severity:
fault_severity=0|1|2

Important
=========
Root Cause, Explanation, and Recommendation targets are structured,
data-constrained synthetic targets because the source dataset contains
fault_severity labels but not engineer-validated root-cause/remediation
annotations. See README.txt.