toxic_LoRA

Toxic Comment Classifier (Mistral 7B + LoRA)

Fine-tuned Mistral 7B using LoRA to classify comments as toxic or not toxic, using the Jigsaw Toxic Comments dataset.

Limitation — Read Before Using

This model has 0% recall on toxic content. It was trained on an imbalanced dataset (4.4% toxic examples) and learned to always predict "not toxic" regardless of input. It will not flag actual toxic content.

This is documented honestly rather than hidden — see Evaluation Results below for the full diagnosis. Do not use this model for actual content moderation in its current state.

Training Details

• Base model: unsloth/mistral-7b-instruct-v0.3-bnb-4bit
• Method: LoRA (r=16, alpha=16), targeting q/k/v/o/gate/up/down projections
• Training data: 2,500 examples from Jigsaw Toxic Comments (111 toxic, 2,389 not toxic)
• Training steps: ~100, batch size 2, gradient accumulation 4

Evaluation Results

Tested on 20 held-out examples (no overlap with training data):

Confusion Matrix:

What I'd Do Differently

1. Balance the training set (equal toxic / not-toxic examples)
2. Evaluate with precision/recall/F1 during training, not just after
3. Check class balance during dataset preparation, before training starts

Why This Is Documented Instead of Hidden

A model reporting fake high accuracy while having zero real capability is worse than no model — it creates false confidence. This card exists so anyone using this model understands its actual (current) limitations.

Intended Use

Currently: Portfolio demonstration only. Not suitable for production content moderation without retraining on a balanced dataset.