Add behavioral_drift: fine-tuning output integrity metric beyond loss curves

[YuhaoLin2005]

What problem does this solve?

Loss curves improve during fine-tuning and everyone ships. But loss can drop while every output degenerates into repeating digit sequences or near-identical strings. Perplexity does not catch this — it measures token-level prediction quality, not output integrity. BLEU/ROUGE do not catch this either — they measure n-gram overlap against ground truth, not behavioral collapse against a baseline.

There is currently no metric in evaluate that catches this class of failure.

How it works

behavioral_drift composites three signals into a single drift_score (0-1, lower = worse):

1. self-BLEU — concatenated-reference BLEU-1 across fine-tuned outputs. High self-BLEU means the model is producing near-identical strings regardless of input (mode collapse).
2. digit density delta — compares the fraction of numeric characters in fine-tuned outputs vs. base model outputs. A sudden spike means the model is producing numeric garbage instead of language.
3. repetition ratio — unique-token / total-token ratio per output. Low ratio means degenerate token looping.

Each signal has a documented threshold. A diagnosis string is returned alongside the score.

Relationship to existing metrics

Not a replacement for perplexity or BLEU. An orthogonal behavioral check. Intended workflow: run drift_score alongside your standard eval suite. If perplexity improves but drift_score drops, you have a silent quality regression.

Implementation

• 147 lines of Python, single file
• Zero external dependencies beyond evaluate + datasets
• Pure deterministic computation — no model calls, no stochastic components
• Demo app (app.py) included for interactive testing via Gradio

[YuhaoLin2005]

Real-world failure case this metric catches

During a 6-round Qwen2.5-1.5B fine-tuning experiment, I observed:

Round	Loss	Actual Behavior
3	9.13	Normal output
4	8.78 (improved!)	Output collapsed to repeating digit sequences

Loss improved. The model broke. No standard metric caught it.

This is the exact failure mode behavioral_drift detects:

• self-BLEU spikes (near-identical outputs regardless of input)
• digit_density_delta spikes (numeric garbage replacing language)
• repetition_ratio drops (degenerate token looping)

The multiplicative formula scores this case as 0.0 (total collapse). An additive alternative gives 0.67 (misleadingly high).

Experiment details and metric design rationale available in the repo: github.com/YuhaoLin2005/training-gate

Happy to add more test cases or adjust threshold defaults per maintainer preference.