---
language: en
license: mit
tags:
  - ci-cd
  - github-actions
  - bert
  - lora
  - peft
  - xai
  - explainable-ai
  - text-classification
  - software-engineering
datasets:
  - facebook/react
metrics:
  - f1
  - accuracy
model-index:
  - name: AdaptCI-XAI
    results:
      - task:
          type: text-classification
        metrics:
          - type: f1
            value: 0.7124
---

# AdaptCI-XAI: CI Pipeline Failure Classifier

**Research project:** IS 8101 — Sabaragamuwa University of Sri Lanka
**Student:** Mary Angel Anton Premathas (20APC4548)
**Paper:** AdaptCI-XAI: Explainable AI for CI Pipeline Failure Diagnosis using Transformer-Based Models on GitHub Actions

## What this model does
Classifies GitHub Actions CI/CD pipeline failure logs into 4 categories:
| Label | Description |
|-------|-------------|
| `config_error` | Malformed YAML, outdated action versions |
| `dependency_failure` | npm/pip install failures, missing packages |
| `test_failure` | Unit/integration test failures, type errors |
| `infrastructure` | Runner timeout, OOM, network errors |

## How to use
```python
from transformers import pipeline
clf = pipeline("text-classification", model="MaryAngel/AdaptCI-XAI")
result = clf("npm ERR ENOENT no such file or directory node_modules/react")
print(result)
# [{'label': 'dependency_failure', 'score': 0.94}]
```

## Training details
- **Base model:** bert-base-uncased
- **Fine-tuning:** LoRA (r=8, alpha=16, target=query+value layers)
- **Trainable parameters:** ~0.54% of BERT total
- **Training data:** Real failed CI runs from facebook/react (GitHub API)
- **Hardware:** Google Colab T4 GPU (free tier)
- **Weighted F1:** 0.7124

## Novelty
1. First LoRA fine-tuning applied to CI/CD log classification
2. First SHAP attribution on CI/CD failure predictions
3. First expertise-aware (novice) adaptive explanation system for CI/CD
4. Multi-source labelling: log text + workflow name + step name signals
5. Fully reproducible on FREE hardware (Colab T4 + HF)

## XAI
Each prediction includes SHAP token attribution showing which log words drove the classification decision — making the black-box model transparent to developers.