| name | vera-ai-nlp-reviewing |
| description | Runs class balance diagnostics and baseline text classification for NLP tasks. Produces class distribution tables, text length statistics, vocabulary analysis, TF-IDF feature inspection, a baseline Logistic Regression classifier with weighted F1 and macro AUC (bootstrapped 95% CIs), confusion matrix, and ROC curves. Ends with a recommendation block listing additional models and analyses available in the analysis workflow. Outputs Python scripts with 2 publication-quality plots. Triggered when user has text data and says "text classification," "NLP," "sentiment analysis," "AI detection," "language model detection," "document classification," "spam detection," "text mining," "topic classification," or describes a task involving classifying text into categories. Does not handle structured/tabular-only data or image data. |
| argument-hint | ["dataset-path-or-description"] |
| allowed-tools | Read, Bash, Write, Edit |
NLP Text Classification --- Data Diagnostics & Baseline Modeling
Table of Contents
Open-source skill.
Scope Boundary
Use this skill when:
- The task is supervised text classification with labeled examples.
- A lightweight, reproducible baseline is the right first checkpoint before deeper models.
Do not use this skill when:
- The task is generation, summarization, retrieval, topic modeling, or sequence labeling.
- Labels are absent, extremely sparse, or only available through prompt-based weak supervision.
- The core signal lives primarily in images, structured covariates, or temporal ordering rather than text.
Workflow
Read each step file in workflow/ before executing that step.
| Step | Responsibility | Executor | Document | Input | Output |
|---|
| Collect | Collect Inputs | Main Agent | workflow/step01-collect-inputs.md | User input | Structured input summary |
| Diagnose | Check Distribution | Main Agent | workflow/step02-check-distribution.md | Prior step output | PART 1 code block |
| Baseline | Run Primary Test | Main Agent | workflow/step03-run-primary-test.md | Prior step output | PART 2-3 code blocks + T1 track artifacts |
Decision Tree
1. CHECK CLASS BALANCE
├── Balanced (minority ≥ 10%) → standard methods
└── Imbalanced (minority < 10%) → class weighting, note power limits
2. TEXT FEATURE STRATEGY
├── Vocab size < 50k → standard TF-IDF (max_features=10000)
└── Vocab size ≥ 50k → sublinear TF-IDF with min_df=5
# 10k features is the baseline default because it keeps sparse linear models
# CPU-friendly while still covering common unigrams/bigrams in small-to-mid NLP datasets.
# The 50k vocabulary gate is a heuristic: once the raw vocabulary gets very large,
# sublinear TF-IDF and min_df filtering usually stabilize the baseline more than
# throwing additional rare terms at Logistic Regression.
3. EXTRA FEATURES
├── Available → text + extra (augmented matrix)
└── Not available → text-only (TF-IDF)
Required Inputs
| Role | What to collect |
|---|
| Text column | Column name containing text data |
| Label column | Target variable, number of classes |
| Group column | For group-aware splitting (optional) |
| Extra features | Numeric columns for augmentation (optional) |
Code Structure
PART 0: Setup & Data Loading
PART 1: Data Diagnostics → plot_01_data_overview.png
PART 2: Baseline Classification → plot_02_confusion_roc.png
PART 3: Recommendation Block → text pointing to analysis workflow
Reporting Standards
- Metrics: weighted F1 and macro AUC (OVR) — always with 95% bootstrapped CIs
- Format: "F1 = 0.XXX, 95% CI [0.XXX, 0.XXX]"
- AUC: "AUC = 0.XXX, 95% CI [0.XXX, 0.XXX]"
- Decimal places: 3 for F1/AUC/p-values, 1 for percentages, 0 for counts
- Proportions: report as percentages with 1 decimal
- Sample size: report final analytic N (train/val/test split sizes)
- Non-significance: "not statistically significant at alpha = .05" — never "no difference"
- Bootstrapped CIs: 1000 iterations, 2.5th/97.5th percentiles
- For imbalanced classes (minority < 10% of total, or per-class < 20%): report Precision–Recall AUC (PR-AUC / Average Precision) alongside F1 and ROC-AUC. PR-AUC is more informative than ROC-AUC when the positive class is rare because it does not reward true-negative inflation. Use
sklearn.metrics.average_precision_score with macro averaging for multi-class. Frame in results as "For class imbalance, PR-AUC is the primary discrimination metric; ROC-AUC is reported for comparability with prior work."
Models Available
| Status | Models |
|---|
| Implemented in this skill | Logistic Regression on TF-IDF features, optionally augmented with extra numeric metadata |
Implemented downstream in vera-ai-nlp-generating | SVM, Random Forest, LightGBM, GRU, TextCNN, ALBERT |
| Not shipped in the open-source build | BERT, RoBERTa, DeBERTa, SetFit, prompt-only classifiers, GPT/Claude few-shot pipelines |
Example Dataset
Any text classification dataset. Default: 20 Newsgroups (sklearn).
Python: from sklearn.datasets import fetch_20newsgroups
Configuration Defaults
Pipeline constants live in config/default.json. Read it before generation — values there override anything in this SKILL.md. Key knobs:
alpha (0.05), ci_level (0.95) — inference thresholds
plot_dpi (300), plot_width, plot_height — figure dimensions
bootstrap_iterations (1000) — bootstrap CIs for metrics
split.{test_size, val_size_of_train, random_state} — data-split reproducibility
tfidf.{max_features, ngram_range, sublinear_tf, min_df} — text featurization
baseline_model.* — LogReg baseline hyperparameter grid
class_balance_threshold (0.10) — minority-class threshold for rare-event warning
To override: create config/local.json; the runtime merges local over default.
Why These Defaults
class_balance_threshold = 0.10 is a warning heuristic, not a theorem. Below about 10% minority prevalence, weighted F1 and ROC-AUC often look healthier than the actual positive-class retrieval story, so the skill promotes PR-AUC in that regime.
bootstrap_iterations = 1000 is the default compromise between stable percentile intervals and a baseline run that still finishes comfortably on CPU.
tfidf.max_features = 10000 keeps the baseline sparse-linear model reproducible and fast enough for smoke tests while still covering the dominant unigram/bigram signal in most benchmark-scale corpora.
Minimal Smoke Test
- Smoke-test prompt: "Run
vera-ai-nlp-reviewing on a 4-class subset of fetch_20newsgroups, using the text as input and the target as label. Produce the standard T1 baseline artifacts."
- Suggested fixture: 200-1000 rows total, 50-250 rows per class, no extra features required.
- Expected pass condition: the run writes one runnable Python script,
methods.md, results.md, tables/, references.bib, and the two baseline figures listed below.
Cross-Skill Interface
Output:
├── code_python → .py script
├── methods_md → methods.md baseline fragment
├── results_md → results.md baseline results fragment
├── tables/ → Markdown/CSV tables for class balance + metrics
├── figures/ → 2 PNGs (class balance + confusion/ROC)
├── references_bib → .bib with baseline/evaluation citations
└── recommendations → text block (what analysis workflow produces)
When this skill is used as T1_baseline inside vera-ai-application-pipelining,
the standardized track artifacts above are REQUIRED, not optional.