A lightweight text classifier for experimental classification of LLM responses as strong or weak.
The project demonstrates a complete small-scale machine learning workflow:
Dataset → feature extraction → model training → cross-validation → model artifact → inference
The classifier is intended as an educational and engineering demonstration rather than a production LLM quality evaluator.
Model description
The classifier uses:
TF-IDF text features
word unigrams and bigrams
logistic regression
balanced class weights
combined prompt and response text as input
The final model is trained on all available examples after cross-validation.
Training data
The model was trained on the public dataset:
MonikaDvorackova/llm-response-evaluation-examples
The training version contains 200 synthetic prompt-response examples:
100 strong
100 weak
10 technical domains
Domains include:
AI governance
LLM reliability
retrieval-augmented generation
MLOps
LLM evaluation
machine learning
data governance
AI safety
NLP
responsible AI
The examples are synthetic and manually designed for experimentation.
They are not production LLM traffic and should not be interpreted as representative of real-world model behaviour.
Evaluation
Two cross-validation strategies were examined.
Stratified 5-fold cross-validation
Random stratified 5-fold cross-validation produced:
Metric
Score
Accuracy
1.000
Macro F1
1.000
Weighted F1
1.000
These results should be interpreted cautiously.
The synthetic dataset contains related response constructions and recurring linguistic patterns. Random stratified splitting can therefore place structurally similar examples in both training and validation folds.
For this reason, the random cross-validation result is likely optimistic and is not treated as evidence of real-world generalization.
Grouped 5-fold cross-validation
A stricter evaluation was performed using the example domain as the grouping variable.
This prevents examples from the same domain from appearing in both the training and validation portions of an individual fold.
The model therefore has to classify examples from held-out domains.
Results:
Metric
Score
Accuracy
0.970
Macro precision
0.972
Macro recall
0.970
Macro F1
0.970
Class-level results:
Class
Precision
Recall
F1
Support
strong
1.000
0.940
0.969
100
weak
0.943
1.000
0.971
100
Confusion matrix:
predicted strong predicted weak
94 6
0 100
The grouped evaluation is considered more informative than the random stratified evaluation for this dataset.
However, it still evaluates generalization only across the synthetic domains represented by the dataset. It does not establish performance on independently collected real-world LLM responses.
Training
The repository contains train.py, which defines the training pipeline and evaluation procedure.
import joblib
model = joblib.load("model.joblib")
prompt = "Why should an ML system be monitored after deployment?"
response = (
"Production conditions may differ from development conditions. "
"Monitoring can help detect data drift, performance degradation, "
"incidents and unexpected behaviour."
)
text = f"PROMPT: {prompt}\nRESPONSE: {response}"
prediction = model.predict([text])[0]
probabilities = model.predict_proba([text])[0]
print("Prediction:", prediction)
print(
dict(
zip(
model.classes_,
probabilities,
)
)
)
Intended use
This model is suitable for:
educational ML experiments
evaluation-pipeline prototyping
testing text-classification workflows
demonstrating reproducible model training
experimenting with synthetic LLM evaluation data
It may also be useful as a simple baseline when comparing deterministic text classifiers with more sophisticated evaluation approaches.
Limitations
This model has substantial limitations.
The training dataset is small and synthetic.
The classifier may learn stylistic or lexical patterns associated with how the examples were constructed rather than general concepts of response quality.
The labels strong and weak are simplified experimental annotations.
The model does not independently determine:
factual correctness
hallucination
safety
regulatory compliance
bias
harmfulness
overall LLM quality
A high classification probability must not be interpreted as a calibrated probability that an LLM response is objectively good or correct.
The reported evaluation results should not be extrapolated to production LLM traffic.
Future work
Possible extensions include:
independently generated evaluation examples
human-annotated responses
larger and more diverse datasets
semantic embeddings
transformer-based classifiers
calibration analysis
adversarial evaluation
pairwise preference data
comparison with LLM-as-a-judge methods
external held-out test sets
Reproducibility
The repository contains:
train.py — training and evaluation pipeline
requirements.txt — Python dependencies
model.joblib — serialized trained model
The associated public dataset documents the data used for training.
